Ann. I. H. Poincaré – AN 24 (2007) 711–739 www.elsevier.com/locate/anihpc
Harnack inequalities, exponential separation, and perturbations of principal Floquet bundles for linear parabolic equations Inégalités de Harnack, séparation exponentielle, et fibré principal de Floquet pour des équations linéaires paraboliques Juraj Húska, Peter Poláˇcik ∗,1 , Mikhail V. Safonov School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA Received 10 June 2005; accepted 13 April 2006 Available online 19 December 2006
Abstract We consider the Dirichlet problem for linear nonautonomous second order parabolic equations with bounded measurable coefficients on bounded Lipschitz domains. Using a new Harnack-type inequality for quotients of positive solutions, we show that each positive solution exponentially dominates any solution which changes sign for all times. We then examine continuity and robustness properties of a principal Floquet bundle and the associated exponential separation under perturbations of the coefficients and the spatial domain. © 2006 Elsevier Masson SAS. All rights reserved. Résumé On considère le problème de Dirichlet pour des équations paraboliques linéaires non autonomes du second ordre avec coefficients bornés mesurables sur un domaine borné de Lipschitz. Utilisant une nouvelle inégalité du type Harnack pour les quotients de solutions strictement positives, on montre que toute solution positive domine exponentiellement toute solution qui change de signe en tout temps. On examine ensuite les propriétés de continuité et de robustesse pour un fibré principal de Floquet et la séparation exponentielle associée à des perturbations des coefficients et du domaine spatial. © 2006 Elsevier Masson SAS. All rights reserved. MSC: 35K20; 35B40; 35B20; 58J35; 35P15 Keywords: Exponential separation; Harnack inequalities; Positive entire solutions; Principal Floquet bundle; Robustness; Perturbations
* Corresponding author.
E-mail addresses:
[email protected] (J. Húska),
[email protected] (P. Poláˇcik),
[email protected] (M.V. Safonov). 1 Supported in part by NSF Grant DMS-0400702.
0294-1449/$ – see front matter © 2006 Elsevier Masson SAS. All rights reserved. doi:10.1016/j.anihpc.2006.04.006
712
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
1. Introduction Consider the following Dirichlet problem for a linear nonautonomous parabolic equation ut + Lu = 0 in Ω × J, u = 0 on ∂Ω × J.
(1.1)
Here Ω ⊂ RN is a bounded domain, J is an open interval in R, and L is a time-dependent second-order elliptic operator of either the divergence form (D) Lu = −∂i aij (x, t)∂j u + ai (x, t)u + bi (x, t)∂i u + c0 (x, t)u or the non-divergence form Lu = −aij (x, t)∂i ∂j u + bi (x, t)∂i u + c0 (x, t)u
(ND)
(we use the summation convention and the notation ∂i = ∂/∂xi ). We assume that the coefficients are real valued and bounded: aij , ai , bi , c0 ∈ Bd0
(i, j = 1, . . . , N ),
(1.2)
where d0 > 0 is a fixed constant and Bd0 := f ∈ L∞ (Ω × R): f L∞ (Ω×R) d0 .
(1.3)
For some results in the case (ND), when explicitly indicated, we in addition assume that aij ∈ C(Ω × R), i, j = 1, . . . , N . We always consider uniformly parabolic equations: there exists α0 > 0 such that aij (x, t)ξi ξj α0 |ξ |2
((x, t) ∈ Ω × R, ξ ∈ RN ).
(1.4) RN .
This means that there are positive Concerning Ω, our assumption is that it is a bounded Lipschitz domain in constants rΩ and mΩ such that for each y ∈ ∂Ω, there is an orthonormal coordinate system centered at y in which (1.5) Ω ∩ BrΩ (y) = x = (x , xN ): x ∈ RN −1 , xN > φ(x ), |x| < rΩ , and ∇φL∞ mΩ .
(1.6) RN
of radius r > 0 and center x. For the remainder of this paper, if not Here and below Br (x) denotes the ball in stated otherwise, we shall assume that Ω is a domain as above with rΩ r0 ,
mΩ M0 ,
diam Ω R0 ,
(1.7)
where r0 , M0 , R0 are fixed positive constants. The main goal of our paper is to examine, in this general context, properties of solutions of (1.1) that are analogous to properties of principal eigenfunctions of time-independent (elliptic) or time-periodic parabolic problems. The analogy is best seen in one of our main results, Theorem 2.6, where we establish the existence of two time-dependent spaces X 1 (t), X 2 (t). For each t ∈ R, these are subspaces of a suitable Banach space X in which the initial conditions
the space of continuous functions vanishing for (1.1) are taken (X = L2 (Ω) in the divergence case and X = C0 (Ω), 1 on the boundary, in the nondivergence case); X (t) is the (one-dimensional) span of a positive function, X 2 (t) \ {0} does not contain any nonnegative function and the subspaces are complementary to one another: X = X 1 (t) ⊕ X 2 (t)
(t ∈ R).
X 2 (s)
is characterized by the property that if a nontrivial solution of (1.1) has its initial condition at The subspace time s contained in X 2 (s), then it changes sign for all t > s. On the other hand, the solutions starting from X = X 1 (s) do not change sign for any t > s. These characterizations imply the invariance of the bundles X i (t), t ∈ R, i = 1, 2: if u1 (t), u2 (t) are solutions with ui (s) ∈ X i (s), then ui (t) ∈ X i (t) for all t > s. Finally, for any such pair of solutions ui , assuming u1 is nontrivial, we have the estimate (which we call the exponential separation) u2 (·, s)X u2 (·, t)X Ce−γ (t−s) u1 (·, t)X u1 (·, s)X
(t s),
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
713
where C, γ > 0 are universal positive constants (they are determined by the quantities N , α0 , d0 , r0 , R0 , M0 from the above assumptions). See Section 2 for the precise formulation of these results. The existence of the invariant bundles with exponential separation extends in a natural way a well known theorem on the existence of a spectral decomposition associated with the principal eigenvalue of elliptic or time-periodic parabolic operators (see [17], for example). Indeed, the principal eigenvalue of a time-independent operator L is simple and has a positive eigenfunction (which spans the first space X 1 in the spectral decomposition) and no other generalized eigenfunction is nonnegative (thus the codimension-one subspace X 2 in the spectral decomposition contains no nontrivial nonnegative function). Moreover, the principal eigenvalue is smaller than the real part of any other eigenvalue, which can be equivalently formulated as exponential separation for the corresponding parabolic equation. Similar results hold for time-periodic parabolic problems in which case one finds a Floquet decomposition X = X 1 (t) ⊕ X 2 (t) with time-periodic spaces. Again X 1 (t) is spanned by a positive function and the solutions in X 2 (t) are exponentially dominated by solutions in X 1 (t). In elliptic and periodic-parabolic problems, theorems on principal eigenvalues and eigenfunctions are classical. In the non-selfadjoint case, they are usually associated with and derived from the Krein–Rutman theorem. A derivation can be found in [17] and many other monographs; for more recent results see [2–4,8]. The origins of the results on invariant bundles and exponential separation in nonautonomous parabolic equations with general time-dependence are in the papers [27,38] (an ODE predecessor is [39]). These results were motivated by certain problems in nonlinear parabolic equations where Eqs. (1.1) naturally appear as linearizations. More specifically, it is well known that if the nonlinear equation has a dissipation property, then each bounded solution approaches a set of entire solutions (that is, solutions defined for all times in R). For further understanding of the behavior of the solution, the linearization of the equation at the entire solutions is often very useful. Now, if the nonlinear equation has a gradient structure, then the entire solutions are steady states and the linearization just gives elliptic operators. However, if no such a priori knowledge of the limit entire solutions is available (this is typically the case in time-periodic equations), to employ the linearization one needs to study Eqs. (1.1) with general time dependence. The principal Floquet bundle and exponential separation then play a similar role as the principal eigenfunction and related spectral properties of elliptic operators do in the local analysis of solutions near steady states. Perhaps the most striking application of such a linearization technique appears in the results on the large time behavior of typical solutions of periodic parabolic problems, as given in [37] and later, with improvements, in [18,41]. For other applications we refer the reader to [18,20,22,28–32,36,37,40,41]. We also remark that in one space dimension, related results on invariant bundles characterized by nodal properties of solutions, as in the classical Sturm–Liouville theory, can be found in [5,42]. See also the survey [35] for a discussion of these results, for both N = 1 and N > 1, and some perspective. Available results on Floquet bundles and exponential separation cover a much broader class of problems than (1.1); in particular, [38,41] deal with abstract parabolic equations admitting a strong comparison principle. However, in the context of the specific problem (1.1), they are restricted by various regularity conditions on the domain and coefficients. The results we present here do not rely on any regularity assumptions other than those mentioned above (continuity of aij in the nondivergence case and Lipschitz continuity of the domain). Another major contribution of our paper is a new approach which leads to more specific results. An example is the fact that the constants C and γ in the exponential separation are determined only by explicit quantities from conditions (1.4), (1.2), and (1.7). We derive the exponential separation result from a new elliptic-type Harnack estimate for quotients of positive solutions of (1.1), which itself has many other interesting consequences regarding solutions of (1.1) (the uniqueness of positive entire solutions among them). A significant portion of this paper is devoted to perturbation results in the divergence case. We prove the continuous dependence of the bundles X 1 (t), X 2 (t), t ∈ R, on the coefficients and the domain and establish a robustness property of the exponential separation. We also prove the continuous dependence of the principal spectrum (see Subsection 2.3 for the definition). These results generalize continuous dependence on the domain and coefficients of the principal eigenvalue and eigenfunctions of uniformly elliptic operators. In this paper we only consider the Dirichlet problem which has many specific features. Indeed, the behavior of the solutions near the boundary ∂Ω, where they vanish, is a major concern in this paper. For the oblique derivative problem, an exponential separation theorem is proved in a similar generality in [19]. The paper is organized as follows. Section 2 contains the statements of our main results. We have grouped them in three subsections. Subsection 2.1 contains the Harnack-type estimate for the quotients of positive solutions and a result on exponential separation between any positive solution and any solution that changes sign for all times. Among corollaries of these theorems, we state the uniqueness of positive entire solutions and a theorem on a universal
714
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
spectral gap for elliptic operators. Subsection 2.2 contains a theorem on invariant bundles and exponential separation, as discussed above, and in Subsection 2.3 we formulate our perturbation results for the divergence case. We have included two preliminary sections; Section 3 contains basic estimates of positive solutions, relying on various Harnack inequalities. In these estimates we make no reference to existence theorems on initial-boundary value problems, thus we do not need the continuity assumption on the coefficients aij in the nondivergence case. In Section 5 we list several properties of the evolution operator associated with Eq. (1.1). The remaining sections consist of the proofs of our theorems. The appendix contains the proof of an existence result from Section 5. 2. Statement of the main results Recall the standing hypothesis that Ω is a bounded Lipschitz domain satisfying (1.7). In the whole section, L is
for the space of continuous as in (D) or (ND) with coefficients satisfying (1.2), (1.4). We use the notation C0 (Ω)
vanishing on ∂Ω. Whenever needed, we assume it is equipped with the supremum norm. functions on Ω We say that a function changes sign if it assumes both positive and negative values. We use the standard notions of solutions of the equation ut + Lu = 0 as well as for the boundary value problem (1.1) and for the initial value problem ut + Lu = 0 in Ω × (s, T ), u = 0 on ∂Ω × (s, T ), u = u0 in Ω × {s},
(2.1)
where s, T ∈ R, s < T . In the divergence case, assuming u0 ∈ L2 (Ω), we consider the usual weak solutions (see [1,24,
× (s, T ) 26]). Under our standing assumptions, (2.1) has a unique weak solution and this solution is continuous on Ω
(more properties of weak solutions are recalled in Sections 3 and 5). In the nondivergence case, we take u0 ∈ C0 (Ω) 2,1,N +1
(Ω × (s, T )) ∩ C(Ω × [s, T ]) and the solution always refers to a strong solution [26], that is, a function u ∈ Wloc which satisfies the initial and boundary conditions everywhere and the equation almost everywhere. The assumed regularity guarantees that the solution is unique. If the coefficients aij , i, j = 1, . . . , N , are continuous on Ω × R then
(see Section 5). the solution exists for each u0 ∈ C0 (Ω) 2.1. Harnack inequality for quotients and exponential separation Our first theorem contains a new Harnack-type estimate for quotients of positive solutions. This is our basic technical tool in the proofs of exponential separation theorems, but it is a result of independent interest. Theorem 2.1. Let δ0 > 0 and let u1 , u2 (u2 ≡ 0) be two nonnegative solutions of (1.1) on Ω × (0, ∞). Then for all t δ0 the following estimate holds sup x∈Ω
u2 (x, t) u2 (x, t) C inf , x∈Ω u1 (x, t) u1 (x, t)
(2.2)
with a constant C > 1 depending only on δ0 , N, α0 , d0 , r0 , R0 , M0 . Theorem 2.1 is an extension of [13, Theorem 4.3] and [12, Theorem 5], where the authors prove boundedness of quotients of positive solutions of (1.1). The next theorem states that sign-changing solutions are exponentially dominated by positive solutions.
T = ∞, and Theorem 2.2. In case L is as in (ND) assume that aij ∈ C(Ω × R), i, j = 1, . . . , N . Let u0 ∈ C0 (Ω), assume that (2.1) has a solution u(·, t; s, u0 ) which changes sign for all t > s. Let moreover v be a positive solution (vanishing on ∂Ω) of the same equation on Ω × (s − 1, ∞). Then there are constants C, γ > 0, depending only on N , α0 , d0 , r0 , R0 , M0 , such that u(·, t; s, u0 )L∞ (Ω) u0 L∞ (Ω) Ce−γ (t−s) v(·, t)L∞ (Ω) v(·, s)L∞ (Ω)
(t s).
(2.3)
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
715
the initial conditions are taken If L is in the divergence form (D), the statement remains valid if instead of C0 (Ω) in Lp (Ω), for some 2 p ∞, and the L∞ -norms in (2.3) are replaced by the Lp -norms. We refer to estimate (2.3) as an exponential separation between sign-changing and positive solutions of (1.1). We emphasize that the constants C and γ in (2.3) depend only on the specific bounds in conditions (1.4), (1.2) and (1.7), and not on the solutions, or directly on L and Ω. We next state several interesting and useful consequences of the previous results. The first one deals with the oscillation in the space domain Ω of the quotient of two solutions of (1.1). Below we shall often abuse the notation slightly and omit the arguments x of the functions considered, for example, we write u(t) for a solution of (1.1). This should cause no confusion. For any continuous function f : Ω × R → R, we define osc f (t) := sup f (x, t) − inf f (x, t), Ω
x∈Ω
x∈Ω
whenever either the supremum or infimum is finite. Corollary 2.3. Let u and v be solutions of (1.1) on Ω × (0, ∞), and let v > 0 on Ω × (0, ∞). Then, assuming the quantities below are defined and finite, we have ω(t) := osc Ω
u(t) ω(s) v(t)
for t s > 0,
(2.4)
and ω(t) μ · ω(s) where
μ := 1 − C −1
(t s + 1)
(2.5)
∈ (0, 1), C > 1 being the constant from Theorem 2.1 with δ0 set equal to 1.
The proof is given in Section 4. It can be verified (see the proof of Lemma 6.1) that the finiteness assumption in Corollary 2.3 is always satisfied in the divergence case and assuming aij ∈ C(Ω × R), i, j = 1, . . . , N , also in the nondivergence case. In the next corollary we state the uniqueness (up to scalar multiples) of positive solutions of (1.1) defined on Ω × R; we refer to such solutions as positive entire solutions. Corollary 2.4. If u1 and u2 are positive entire solutions of (1.1), then there is a constant q such that u2 ≡ qu1 . For nonautonomous parabolic equations in the divergence form, under additional conditions on the coefficients and the domain, the uniqueness of entire solutions was proved in earlier papers; the first one seems to be that of Nishio [34], later results can be found in [29,30,36]. We gave a simple proof of Corollary 2.4 for equations in divergence form in [20]. The approach we used there originated in [36]. Our present proof, see Section 4, is different, still rather simple, and, being based on Theorem 2.1, it works for both types of equations. The fact that the constants appearing in the estimates of Theorems 2.1, 2.2 depend only on the bounds in (1.4), (1.2), and (1.7), and not directly on L, has an interesting consequence for elliptic operators. Namely, it allows us to establish a universal gap between the first (principal) eigenvalue and the rest of the spectrum. This result is nontrivial for non-divergence form operators and to our knowledge it has not been noted so far (a proof for operators without lower order terms is given in [21]). With L independent of t, consider the eigenvalue problem Lu = λu in Ω, u = 0 on ∂Ω.
(2.6)
The principal eigenvalue λ1 of this problem is the eigenvalue which is real and has a positive eigenfunction. It is well known (see for example [3,10]) that λ1 is well defined by these requirements. Corollary 2.5. Assume the coefficients of L are independent of t. In case L is as in (ND) assume that aij ∈ C(Ω), i, j = 1, . . . , N . Let λ1 be the principal eigenvalue and let λ be any other eigenvalue of (2.6). Then Re(λ) − λ1 γ > 0, where γ = γ (N, α0 , d0 , r0 , R0 , M0 ) is as in (2.3).
(2.7)
716
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
Proof. Let ϕ1 > 0 be an eigenfunction associated with λ1 . The function v1 (x, t) := e−λ1 t ϕ1 (x) is a positive entire solution of (1.1). Let λ = λ1 , λ = a + ib (a, b ∈ R), be an eigenvalue of (2.6) and let ϕλ , ϕλ (x) = f (x) + ig(x), x ∈ Ω, denote an eigenfunction corresponding to λ. For (x, t) ∈ Ω × R define u1 (x, t) := Re e−λt ϕλ (x) = e−at cos(bt)f (x) + sin(bt)g(x) . This function is also an entire solution of (1.1). Since eat u1 (·, t) is periodic in t, the maximum principle implies that u1 (·, t) is either of one sign for t ∈ R or it changes sign for all t ∈ R. Assume u1 (·, t) is, say, positive for t ∈ R. In view of λ = λ1 , the functions u1 , v1 are then two linearly independent positive entire solutions of (1.1), violating Corollary 2.4. Thus u1 (·, t) changes sign for all t ∈ R and applying Theorem 2.2, we conclude that a − λ1 γ , with γ as in (2.3). 2 2.2. Exponential separation and Floquet bundles In the following result, we state the conclusion of Theorem 2.2 in a different form and complement it by additional information, introducing the principal Floquet bundles. We denote by u(·, t; s, u0 ) the solution of (1.1) with the initial
condition u(·, s) = u0 ∈ X. Here X stands for L2 (Ω) with the standard norm in case L is as in (D) and X = C0 (Ω) if L is as in (ND). The existence and uniqueness theorems for the initial-boundary value problems are recalled in Section 5.
× R → R define For any continuous f : Ω λ(f ) = lim inf
log f (·, t)X − log f (·, s)X , t −s
(2.8)
¯ ) = lim sup λ(f
log f (·, t)X − log f (·, s)X . t −s
(2.9)
t−s→∞
and t−s→∞
Theorem 2.6. In case L is as in (ND) assume that aij ∈ C(Ω × R), i, j = 1, . . . , N . The following statements hold true. (i) There exists a unique positive entire solution ϕL of (1.1) satisfying ϕL (·, 0)L2 (Ω) = 1. This solution satisfies −C λ(ϕL ) λ¯ (ϕL ) C for some positive constant C = C(N, α0 , d0 , r0 , R0 , M0 ). (ii) Set XL1 (t) := span ϕL (·, t) , XL2 (t) := u0 ∈ X: u(·, t˜, t, u0 ) has a zero in Ω for all t˜ > t . These sets are closed subspaces of X. They are invariant under (1.1) in the following sense: if i ∈ {1, 2}, u0 ∈ XLi (s), then u(·, t; s, u0 ) ∈ XLi (t) (t s). Moreover, XL1 (t), XL2 (t) are complementary subspaces of X: X = XL1 (t) ⊕ XL2 (t)
(t ∈ R).
(2.10)
(iii) There are constants C, γ > 0, depending only on N , α0 , d0 , r0 , R0 , M0 , such that for any u0 ∈ XL2 (s) one has u0 X u(·, t; s, u0 )X Ce−γ (t−s) ϕL (·, t)X ϕL (·, s)X
(t s).
(2.11)
We refer to the collection of the one-dimensional spaces XL1 (t), t ∈ R, as the principal Floquet bundle of (1.1) and to XL2 (t), t ∈ R, as its complementary Floquet bundle. Property (iii) as stated is an exponential separation between these two bundles. Following [31,32], we call the interval [λ(ϕL ), λ¯ (ϕL )] the principal spectrum of (1.1). As discussed in the introduction, the existence of the Floquet bundles with exponential separation extends in a natural way results on spectral decomposition associated with the principal eigenvalue of time-independent or time-periodic parabolic problems. The principal spectrum [λ(ϕL ), λ¯ (ϕL )] and the positive entire solution ϕL serve as analogues of the principal eigenvalue and eigenfunction.
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
717
2.3. Perturbation results in the divergence case In this subsection we assume that L is in the divergence form (D). In this case, the invariant bundle XL2 (t), t ∈ R, as introduced in the previous section, can be characterized using the principal Floquet bundle for the adjoint equation to (1.1). This way we are also able to study continuity properties of both bundles XL1 (t), XL2 (t) under perturbations of the coefficients and the domain. We also examine the robustness properties of the associated exponential separation and continuity properties of the principal spectrum. Let us first introduce the adjoint problem to (1.1): − vt + L∗ v = 0 in Ω × J, v = 0 on ∂Ω × J,
(2.12)
with L∗ defined by L∗ v = −∂j aij (x, t)∂i v + bj (x, t)v + aj (x, t)∂j v + c0 (x, t)v.
(2.13)
The next theorem characterizes the invariant bundle XL2 (t), t ∈ R, using an entire solution of (2.12). Theorem 2.7. Assume L is as in (D). There exists a unique positive entire solution ψL of (2.12), satisfying ¯ L ) C for some positive constant C = C(N, α0 , d0 , r0 , R0 , M0 ). ψL (0)L2 (Ω) = 1. It satisfies −C λ(ψL ) λ(ψ 2 The space XL (t) defined in Theorem 2.6 can be characterized as 2 2 XL (t) := v ∈ L (Ω): ψL (x, t)v(x) dx = 0 (t ∈ R). Ω
In our first perturbation result, we assume that the domain Ω in (1.1) is fixed. Given an operator L, we say that L admits exponential separation with bound C and exponent γ if for the entire solution ϕL and the invariant bundle XL2 (t), t ∈ R, inequality (2.11) holds for each u0 ∈ XL2 (s). Theorem 2.6 guarantees the existence of a bound and exponent which are common to all operators satisfying (1.2) and (1.4), but now we are interested in robustness of the exponential separation with a possibly larger specific exponent for L. Theorem 2.8. Assume L is as in (D) and let L˜ be another operator of the form (D) with coefficients (a˜ ij )N i,j =1 , ˜ a˜ i , bi , c˜0 , i = 1, . . . , N , satisfying (1.2) and (1.4). Assume that for some δ > 0 aij − a˜ ij , ai − a˜ i , bi − b˜i , c0 − c˜0 ⊆ [0, δ] (i, j = 1, . . . , N ), (2.14) where · is the norm of L∞ (Ω × R). Then the following statements hold true. (i) For each ε > 0 there exist a constant δ1 > 0, depending only on ε and N , α0 , d0 , r0 , R0 , M0 , such that if δ δ1 then ¯ ˜ ) ε (2.15) max λ(ϕL ) − λ(ϕ ˜ ) , ¯λ(ϕL ) − λ(ϕ L
and
L
ϕL˜ (t) ϕL (t)
− ε
ϕ (t) ∞ ϕL˜ (t)L∞ (Ω) L∞ (Ω) L L (Ω)
(t ∈ R).
(2.16)
The statement is also valid with ϕL and ϕL˜ replaced by ψL and ψL˜ , respectively. (ii) Suppose that L admits exponential separation with bound CL and exponent γL . For each ε > 0 there exists δ2 depending only on ε, CL , γL , N , α0 , d0 , r0 , R0 , M0 , such that if δ δ2 then L˜ admits exponential separation with some bound C(ε, CL , γL ) > 0 and exponent γL˜ γL − ε. Remark 2.9. If the L∞ -norms in (2.16) are replaced by the L2 -norms, the following more precise result can be proved (see Section 8). There exist positive constants δ0 and C depending only on N , α0 , d0 , r0 , R0 , M0 , such that if (2.14) holds with δ δ0 , then
ϕL˜ (t) ϕL (t)
− Cδ (t ∈ R). (2.17)
ϕ (t) 2 ϕL˜ (t)L2 (Ω) L2 (Ω) L L (Ω)
718
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
In our last theorem we assume that L is fixed and we vary the domain Ω. We use the phrase that Ω admits exponential separation with bound C and exponent γ , in an analogous way as when varying L above. For each bounded Lipschitz domain Ω denote by ϕΩ , ψΩ the unique positive entire solutions of (1.1), (2.12), respectively we denote the Hausdorff distance of ∂Ω and ∂ Ω: satisfying ϕΩ (0)L2 (Ω) = ψΩ (0)L2 (Ω) = 1. By d(∂Ω, ∂ Ω) = max dist(∂Ω, ∂ Ω), dist(∂ Ω, ∂Ω) , d(∂Ω, ∂ Ω) = sup dist(x, ∂ Ω). dist(∂Ω, ∂ Ω) x∈∂Ω
Theorem 2.10. Assume L is a fixed operator as in (D) with coefficients defined on RN +1 and satisfying the ellipticity be two Lipschitz domains such that their Lipschitz and boundedness conditions (1.2), (1.4) on RN +1 . Let Ω, Ω constants satisfy (1.7). Then the following statements hold. δ1 , then (i) For each ε > 0 there exists δ1 > 0 depending only on ε, N, α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω, ∂ Ω) ¯ Ω ) − λ(ϕ ¯ Ω ) ε (2.18) max λ(ϕΩ ) − λ(ϕΩ ) , λ(ϕ and
ϕΩ (t) ϕΩ (t)
− ε
ϕ (t) ∞
ϕΩ (t)L∞ (Ω) L∞ (RN ) Ω L (Ω)
(t ∈ R).
(2.19)
Estimates (2.18) and (2.19) are also valid with ϕΩ and ϕΩ replaced by ψΩ and ψΩ , respectively. (ii) Suppose that Ω admits exponential separation with bound CΩ and exponent γΩ . Then for each ε > 0 there δ2 , then Ω admits exists δ2 > 0 depending only on ε, CΩ , γΩ , N, α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω, ∂ Ω) exponential separation with some bound C(ε, CΩ , γΩ ) > 0 and exponent γΩ γΩ − ε. Remark 2.11. In estimate (2.19) the functions ϕΩ (·, t) and ϕΩ (·, t), t ∈ R, are thought of as extended by zero outside respectively. Ω and Ω, Theorems 2.8 and 2.10 give robustness of the exponential separation and continuity of the principal spectrum and principal Floquet bundle under perturbations of the operator L and the domain Ω. More precisely, (2.15) and (2.18) imply that the principal spectrum depends continuously on such perturbations. Similarly, (2.17), (2.19) give the continuous dependence of the principal Floquet bundles. The interesting and nontrivial part here is that in our general time-dependent case the estimates are uniform with respect to t ∈ R (cp. [20, Theorem 1.1(i)]). These results extend the well-known continuity properties of the principal eigenvalue and eigenfunction of elliptic and time-periodic parabolic operators (see [2,3,6–8], for example). It is also well-known that in the elliptic or time-periodic case, the gap between the principal eigenvalue and the rest of the spectrum of the corresponding operator is a lower semicontinuous function of the domain and the coefficients. Our lower estimate on γL˜ (γΩ ) extends this result to general timedependent equations. The robustness results do not seem to hold in the non-divergence case in the same generality. The main obstacle is the lack of similar continuity properties of the evolution operator as in the divergence case, see Sections 8, 9. 3. Preliminaries I: Harnack inequalities and estimates of positive solutions For the results in this section we do not need any existence theorem for (1.1). In particular, no continuity assumption on the coefficients of L are needed. For simplicity of notation we state the results on the interval (0, ∞), the results being true on any interval. First we state the comparison principle (see [26]). Theorem 3.1. Let u1 , u2 be two solutions of (1.1) on Ω × (0, ∞), and let u1 u2 on Ω × {0}. Then u1 u2 on Ω × (0, ∞).
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
719
Corollary 3.2. Let u1 , u2 be two solutions of (1.1) on Ω × (0, ∞), and let u1 > 0 on Ω × (0, ∞). Then we have for any s > 0 u2 u2 u2 u2 = M(s) := sup , inf = m(s) := inf . (3.1) sup Ω×{s} Ω×(s,∞) u u u u1 1 Ω×{s} 1 Ω×(s,∞) 1 In particular, M(t) M(s),
m(t) m(s)
for t s > 0.
(3.2)
Proof. The inequality for M(t) is trivial if M(s) = ∞, and it follows directly from Theorem 3.1 if M(s) = 0 or M(s) = 1. In case M(s) ∈ / {0, ∞}, dividing u2 by M(s), we can reduce the proof of the first equality in (3.1) to the case M(s) = 1. The equality for m(t) follows by changing the sign of u2 . 2 In case L is as in (ND), the following theorem is a simple consequence of the standard maximum principle (see e.g. [26, Theorem 7.1]). In case L is as in (D), it is proved in [24, Theorem III.7.1].
× [0, ∞)). Then there is a constant C Theorem 3.3. Let u be a solution of (1.1) on Ω × (0, ∞) such that u ∈ C(Ω depending on N , α0 , and d0 , such that for any t 0 one has
u(t + τ ) ∞ C u(t) L∞ (Ω) (τ ∈ [0, 1]). (3.3) L (Ω) Next we state the interior Harnack inequality [23,26,33]. For any δ > 0, define Ω δ := x ∈ Ω: dist(x, ∂Ω) > δ . Theorem 3.4. Let v be a nonnegative solution of vt + Lv = 0 on Ω × (0, ∞) and let T > 0 be arbitrary. Suppose δ ∈ (0, T ) is such that δ r0 /2 (r0 as in (1.7)) and let X = (x, t), Y = (y, s) be such that x, y ∈ Ω δ , s δ 2 , and T t − s δ 2 . Then there is a positive constant C depending only on δ, T , N , α0 , d0 , R0 such that one has v(Y ) Cv(X).
(3.4)
To state the next results, we introduce some notation. For X = (x, t) ∈ RN +1 , we define a parabolic cylinder to be Cr (X) = Cr (x, t) ≡ Br (x) × t − r 2 , t + r 2 . Further, let us denote Qr (X) = Ω × (0, ∞) ∩ Cr (X). According to the Lipschitz properties of Ω, for y ∈ ∂Ω there is an orthonormal system with y as the origin (0, 0) and (0, r) ∈ Ω for all r ∈ (0, r0 ]. The new coordinates for Y = (y, s) ∈ RN +1 are (0, 0, s). In this coordinate system write Y r = 0, r, s − 2r 2 . Y r = 0, r, s + 2r 2 , We now recall two results from [12,13]. The first one is often referred to as a boundary Harnack inequality. We say that a function defined on an open set Q ⊂ RN +1 continuously vanishes on Γ ⊂ ∂Q if it has a continuous extension to Q ∪ Γ , which vanishes on Γ . √ Theorem 3.5. Let Y = (y, s) ∈ ∂Ω × (0, ∞) and 0 < r 12 min(r0 , s ). Then for any nonnegative solution v of vt + Lv = 0 on Ω × (0, ∞) which continuously vanishes on (∂Ω × (0, ∞)) ∩ C2r (Y ), we have sup v Cv(Y r ). Qr (Y )
The constant C depends only on N , d0 , α0 , M0 . The next estimate states that the quotient of two positive solutions is bounded near the portion of the lateral boundary where each solution vanishes.
720
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
Theorem 3.6. Let X0 = (x0 , t0 ) ∈ ∂Ω × (0, ∞). Assume that u and v are two positive solutions √ of wt + Lw = 0 on Ω × (0, ∞) which continuously vanish on (∂Ω × (0, ∞)) ∩ C2r (X0 ) with 0 < r 12 min(r0 , t0 ). Then u(X0r ) u C . v(X0 r ) Qr (X0 ) v sup
The constant C depends only on N , d0 , α0 , M0 . Theorems 3.5, 3.6 are proved in [12] for equations of divergence form and in [13] for equations of nondivergence form. Although only equations without lower order terms are considered in these papers, the results hold also in the general case. Indeed, since these results are local (solutions are required to vanish on a portion of the boundary only), they carry over to the present situation by considering the method of additional variable. This is done in two steps. First, one assumes that the coefficients of the equation in question are smooth. Applying the method of additional variable (see, e.g., [14] or [20] for details), one derives the results for equations containing lower order terms. Finally, since all the estimates are independent of smoothness, one can take limits (as in [14]) to obtain the results in our more general setting. Combining the above results, we will be able to prove the following elliptic-type Harnack inequality. √ Theorem 3.7. Suppose v is a nonnegative solution of (1.1) on Ω ×(0, ∞) and let T ∈ (0, ∞), 0 < r 12 min(r0 , T ). Then sup Ω r ×(r 2 ,T )
vC
inf
Ω r ×(r 2 ,T )
(3.5)
v.
The constant C depends only on r, T , N , α0 , d0 , M0 , R0 . Let us mention that this inequality is known for equations without lower order terms (see [11,15]). The general case follows from the above results as in the proof of Theorem 1.3 in [11]. Since the proof is short we include it here for the reader’s convenience (note that the method of additional variable does not apply here as the result is not local). Proof of Theorem 3.7. Let (x0 , t0 ), (x1 , t1 ) ∈ Ω r × (r 2 , T ). Theorems 3.5 and 3.4 imply that for all x2 ∈ Ω such that dist(x2 , ∂Ω) r one has supx∈Ω v(x, r 2 /8) Cv(x2 , r 2 /4). Applying (3.4) to the right-hand side of the last inequality and using Theorem 3.3, we obtain for i = 0, 1 sup v x, r 2 /8 C1 v(xi , ti ) C1 sup v(x, ti ) C2 sup v x, r 2 /8 , (3.6) x∈Ω
x∈Ω
x∈Ω
with a constant C2 depending only on r, T , N , α0 , d0 , M0 , R0 . Inequality (3.6) with i = 0, 1 implies (3.5).
2
The next proposition will be used in the proof of Lemma 3.9 below. It is proved in [12,13]. Proposition 3.8. Let √ v be a nonnegative solution of vt + Lv = 0 on Ω × (0, ∞). Take Y = (y, s) ∈ ∂Ω × (0, ∞) and 0 < r 12 min(r0 , s ). Then v(Y r ) Cr θ inf d −θ v , Qr (Y )
where d = d(x) ≡ dist(x, ∂Ω), and C, θ are positive constants depending only on N , d0 , α0 , r0 , R0 , M0 . The following lemma contains an important pointwise estimate, which is also of independent interest. In a slightly weaker form the lemma has been proved in [20]. Lemma 3.9. For any δ > 0 and any positive solution v of (1.1) on Ω × (0, ∞) one has θ v(x, t) (x, t) ∈ Ω × [δ, ∞) , C d(x) ∞ v(t)L (Ω)
(3.7)
where d(x) = dist(x, ∂Ω) and θ is as in Proposition 3.8. The constant C depends only on δ, N , d0 , α0 , r0 , R0 , M0 .
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
721
√ Proof. Fix t δ, r1 = 12 min(r0 , δ). Note that if x ∈ Ω is such that d(x) = dist(x, ∂Ω) r1 / M02 + 1, then (3.6) implies (3.7) with a constant C depending only on δ, N , d0 , α0 , r0 , R0 , M0 . We now consider x ∈ Ω such that d(x)
r1 / M02 + 1. Using the notation introduced just before Theorem 3.5, we let y = (0, 0) ∈ ∂Ω. An easy geometric argument shows that for r ∈ (0, r0 ) we have
r dist (0, r), ∂Ω ∈ ,r . (3.8) M02 + 1
Our assumption on d(x) implies that there exists Y = (y, t) ∈ ∂Ω × {t} such that (x, t) ∈ Qr1 (Y ). Moreover, we can find x˜ ∈ Ω such that Y r1 = (x, ˜ t − 2r12 ). Using Proposition 3.8, we see that ˜ t − 2r12 . v(x, t) C −1 r1−θ d(x)θ v x, Notice that by (3.8) we have d(x) ˜ ∈ [r1 / M02 + 1, r1 ]. Thus, by what we said at the beginning of this proof, we have C1 v(x, ˜ t − 2r12 ) v(t − 2r12 )L∞ (Ω) for some positive constant C1 depending only on r1 , N , d0 , α0 , r0 , R0 , M0 . Finally, by Theorem 3.3, C2 v(t − 2r12 )L∞ (Ω) v(t)L∞ (Ω) . Combining the above estimates we obtain the desired inequality with C = C −1 C1 C2 . 2 As an immediate consequence we get Corollary 3.10. Let δ > 0. Then there exists a positive constant C depending only on δ, N , d0 , α0 , r0 , R0 , M0 , such that if v is a positive solution of (1.1) on Ω × (0, ∞) one has for any τ ∈ [0, 1]
1 v(t + τ )L∞ (Ω) ,C (t > δ). (3.9) ∈ v(t)L∞ (Ω) C Proof. The upper bound is a trivial consequence of Theorem 3.3. The lower bound is proved as follows. Choose a point x0 ∈ Ω such that dist(x0 , ∂Ω) r0 /(1 + M0 ). The Harnack inequality (3.4) implies v(x0 , t) Cv(x0 , t + 1) with some positive constant C. Our special choice of x0 and Lemma 3.9 imply that v(x0 , t + i) is comparable to v(t + i)L∞ (Ω) , i = 0, 1, and we thus get the desired lower bound for τ = 1. Combining it with Theorem 3.3 we get the lower bound for any τ ∈ [0, 1]. 2 4. Proofs of Theorem 2.1 and its corollaries
Proof of Theorem 2.1. The desired estimate will be obtained by combining the results from the previous section. We start by a preliminary estimate. Fix an arbitrary X0 = (x0 , t0 ) with x0 ∈ ∂Ω and t0 δ0 . Set √ r1 δ0 r0 r1 := min , , ρ := , 2 2 1 + M0 Qr1 := Ω ∩ Br1 (x0 ) × t0 − r12 , t0 + r12 . By Theorem 3.6, sup Qr1
u2 u2 (X0 r1 ) C1 u1 u1 (X0 r )
(4.1)
1
with a constant C1 1 depending only on N , d0 , α0 , M0 . As in the proof of Lemma 3.9 we can write X0 r = 1
(y0 , t0 − 2r12 ) for some y0 ∈ Ω ∩ Br1 (x0 ), such that d(y0 ) := dist(y0 , ∂Ω) > ρ. Note that the points (y0 , t0 ± 2r12 ) belong to the cylinder δ0 δ0 ρ ρ Q := Ω × t0 − , t0 + (4.2) , where Ω ρ := x ∈ Ω: d(x) > ρ . 4 4
722
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
Therefore, the previous estimate implies u2 sup C1 M2 m−1 uj , mj := infρ uj . 1 , where Mj := sup Q u Qr1 1 Qρ
(4.3)
We now derive inequality (2.2) for any t δ0 . Take t0 = t. For each x ∈ Ω consider separately the two possible cases: (i) d(x) := dist(x, ∂Ω) ρ, and (ii) d(x) > ρ. In the case (i), we take x0 ∈ ∂Ω such that |x − x0 | = d(x) ρ < r1 , so ρ that (x, t0 ) ∈ Qr1 . Then, by (4.3), u2 (x, t0 )/u1 (x, t0 ) does not exceed C1 M2 m−1 1 . In the other case (ii), (x, t0 ) ∈ Q , −1 so that u2 (x, t0 )/u1 (x, t0 ) M2 m−1 1 C1 M2 m1 as well. We have thus shown that u2 C1 M2 . m1 Ω×{t} u1
M(t) := sup
(4.4)
Interchanging u1 and u2 , we also get u2 u1 −1 C1 M1 −1 m2 = sup = . m(t) := inf Ω×{t} u1 u m C 2 1 M1 Ω×{t} 2
(4.5)
Now it remains to apply (3.5) which guarantees that Mj C2 mj for j = 1, 2, with another constant C2 1 depending only on δ0 , N, α0 , d0 , r0 , R0 , M0 . This gives the desired estimate (2.2) with C := C12 C22 . 2 Proof of Corollary 2.3. Using Corollary 3.2 with u1 := v and u2 := u, we immediately get (2.4): ω(t) = M(t) − m(t) M(s) − m(s) = ω(s)
(t s).
For the proof of (2.5), we use Corollary 3.2 with u1 := v and u2 (x, t) := u(x, t) − m0 (s) · u1 (x, t),
where m0 (s) := inf
Ω×{s}
u . u1
(Note that we are assuming that m0 (s) is finite.) Since m(s) = 0, the corollary gives u2 0 in Ω × (s, ∞), and (2.5) then follows from Theorem 2.1: ω(t) = M(t) − m(t) 1 − C −1 M(t) = μM(t) μM(s) = μ M(s) − m(s) = μω(s) (t s + 1). Thus the corollary is proved.
2
Proof of Corollary 2.4. For any t ∈ R define u2 (t) u2 (t) max (t) := sup , min (t) := inf . Ω u1 (t) u1 (t) Ω Thus oscΩ (u2 (t)/u1 (t)) = max (t) − min (t). Theorem 2.1 implies that oscΩ (u2 (t)/u1 (t)) < ∞ for all t ∈ R. By Corollary 3.1, max (t) is nonincreasing and min (t) is nondecreasing. By Corollary 2.3, we have u2 (t) = 0. lim oscΩ t→∞ u1 (t) Thus for some q ∈ (0, ∞) lim max (t) = lim min (t) = q.
t→∞
t→∞
It follows that the function u = u2 − qu1 is a solution of (1.1) on Ω × R, which vanishes somewhere in Ω for all t ∈ R and consequently infΩ (u2 (t)/qu1 (t)) 1 for all t. Using this fact and Theorem 2.1 we discover that sup Ω
u2 (t) u2 (t) C inf C, Ω qu1 (t) qu1 (t)
with a constant C independent of t ∈ R. This implies that oscΩ (u2 (t)/qu1 (t)) is bounded on all of R. Moreover, by Corollary 2.3, this function is exponentially decreasing on R. This is only possible if oscΩ (u2 (t)/qu1 (t)) ≡ 0 for all t ∈ R and that happens only if u2 ≡ qu1 . The proof is complete. 2
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
723
5. Preliminaries II: The evolution operator Consider the following initial value problem ut + Lu = 0 in Ω × (s, T ), u = 0 on ∂Ω × (s, T ), u = u0 in Ω × {s},
(5.1)
where s, T ∈ R, s < T . In this section we give the basic existence result for this problem and list several properties of the corresponding evolution operator. Since the assumptions differ slightly in the divergence and non-divergence cases, we treat these cases separately. 5.1. Divergence case In this subsection we assume that L is as in (D) with coefficients satisfying (1.4) and (1.2). Recall that our notion of the solution of (5.1) coincides with the notion of the weak solution from [24,26]. For any u0 ∈ L2 (Ω), and s < T , the weak solution of (5.1) exists, it is unique and can be (uniquely) extended to a solution on (s, ∞). Denote the solution by U (t, s)u0 , t s. Let · p,q stand for the operator norm of the space L(Lp (Ω), Lq (Ω)) of bounded linear operators from Lp (Ω) to Lq (Ω). It is well known that the evolution operator U (t, s), t s, satisfies the following Lp − Lq estimates (see [9], for example). Proposition 5.1. For all 1 p q ∞, t, s ∈ R, t > s, one has U (t, s) ∈ L(Lp (Ω), Lq (Ω)) and
− N2 ( p1 − q1 ) ω(t−s)
U (t, s) p e , L(L (Ω),Lq (Ω)) C(t − s) where C 1 and ω ∈ R are constants depending only on N , d0 , α0 . Moreover, for any u0 ∈ L2 (Ω) and T s one has U (·, s)u0 ∈ C([s, T ]; L2 (Ω)). Another property of the evolution operator U (t, s) we will use is positivity. For any p ∈ (1, ∞) and u0 ∈ Lp (Ω), u0 0, we have U (t, s)u0 0 for all t s (see [9]). This can be improved on: nonnegative nontrivial solutions are strictly positive. It is a direct consequence of the Harnack inequality (3.4). Besides positivity, the evolution operator has a smoothing property. In this regard, we mention the following standard regularity result [24, Chapter III, Theorem 10.1]. We use the usual notation for the parabolic Hölder spaces. α
α,
× Theorem 5.2. Let u be the (weak) solution of (5.1) with u0 ∈ L2 (Ω). Then for any T > s we have u ∈ Cloc2 (Ω α (s, T ]), and for any δ > 0 (δ < T − s) the norm u α, 2
is estimated from above by a constant dependC
(Ω×[s+δ,T ])
ing only on N , d0 in (1.2), supΩ×(s,T ) |u|, α0 in (1.4), constants r0 , R0 , M0 in (1.7) and δ. The exponent α > 0 is determined only by N, d0 , α0 .
∩ C0 (Ω)
for some β > 0, then the norm u α, α If, in addition, it is known that u0 ∈ C β (Ω) is estimated
C 2 (Ω×[s,T ]) from above by a constant depending only on N , d0 , supΩ×(s,T ) |u|, α0 , r0 , R0 , M0 , β and the norm u0 C β (Ω)
. The exponent α belongs to (0, β] and is determined by N, d0 , α0 . Combining Lp –Lq estimates, Corollary 3.10, and Lemma 3.9, one proves the following result. Corollary 5.3. There exists a constant C depending only on N , d0 , α0 , r0 , R0 , M0 , such that for any 1 p, q ∞ and any positive entire solution u of (1.1) the following statements hold: (i)
θ u(x, t) C d(x) u(t)Lp (Ω)
((x, t) ∈ Ω × (−∞, ∞)),
(5.2)
where d(x) = dist(x, ∂Ω) and θ is as in Lemma 3.9 (it is a constant depending only on N , d0 , α0 , r0 , R0 and M0 ).
724
(ii)
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
sup
|t−s|1 s,t∈R
u(s)Lp (Ω) C. u(t)Lq (Ω)
(5.3)
Let us now turn our attention to (2.12). One can prove (see [9]) that there is a well defined evolution operator, henceforth denoted by U ∗ (t, s), t s, for the adjoint problem (2.12). Reversing time, we obtain (−t, −s) U ∗ (t, s) = U
(t s),
(t, s), t s, is the (“forward”) evolution operator for the problem where U wt + L˜ ∗ w = 0,
x ∈ Ω,
w = 0 x ∈ ∂Ω, where L˜ ∗ is obtained from L∗ by replacing t with −t. This problem is of the same form as (5.1) and thus U ∗ (t, s) has the same smoothing and positivity properties as U (t, s). We will use this fact frequently without notice. The following proposition summarizes some properties of weak Green’s functions we will need later. The first three statements below are proved in [1] and for the last one we refer the reader to [9]. Proposition 5.4. There exists a unique weak Green’s function k(x, t; ξ, s) associated with (5.1) with the following properties. (i) For any u0 ∈ L2 (Ω) the solution u = U (t, s)u0 of (5.1) is given by u(x, t) = k(x, t; ξ, s)u0 (ξ ) dξ ((x, t) ∈ Ω × (s, ∞)), Ω
and the solution v = U ∗ (s, t)u0 of the adjoint problem (2.12) is given by v(ξ, s) = k(x, t; ξ, s)u0 (x) dx ((ξ, s) ∈ Ω × (−∞, t)). Ω
(ii) The function k(·, ·; ξ, s): Ω × (s, ∞) → R is a positive solution of (1.1) on Ω × (s, ∞). (iii) The function k(·, ·; ξ, s) is bounded on Ω × (s + τ, s + 1) by a constant C(τ ) depending only on τ ∈ (0, 1), N , α0 and d0 . (iv) Let Ω1 ⊆ Ω2 be two bounded domains in RN and let ki (x, t; ξ, s), i = 1, 2, be the corresponding weak Green’s functions. Then, extending k1 (·, ·; ξ, s) by zero outside Ω1 , we have k1 (x, t; ξ, s) k2 (x, t; ξ, s)
((x, t) ∈ Ω2 × (s, ∞)).
The integral representation of solutions, as given in statement (i), and the Fubini theorem readily imply the following identity (5.4) U (t, s)u, v = u, U ∗ (s, t)v (u, v ∈ L2 (Ω), t s). 5.2. Non-divergence case Let us now handle the case when L is as in (ND). In this case we assume, in addition to (1.4) and (1.2), that aij ∈ C(Ω × R), i, j = 1, . . . , N . Consider again the initial value problem (5.1). Recall that this time we consider strong solutions as defined in the beginning of Section 2. We have the following result regarding the well-posedness.
Then the initial value problem (5.1) has a unique solution u; the solution is Proposition 5.5. Assume u0 ∈ C0 (Ω). 2,1,p
× [s, T ]) for all p > 1 and it satisfies the following estimate contained in Wloc (Ω × (s, T )) ∩ C(Ω
u(t) ∞ em(t−s) u0 L∞ (Ω) (t ∈ [s, T ]), (5.5) L (Ω) where m = supΩ×(s,T ) (−c0 ) d0 (d0 is as in (1.2)).
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
725
This result can be derived from uniform estimates of the moduli of continuity of solutions vanishing on the lateral side of a Lipschitz cylinder. In a more general setting, these estimates are contained in Theorem 6.32 (in the case (D)) and Corollary 7.30 (in the case (ND)) of the monograph [26]. For completeness of presentation, we give an alternative proof in Appendix A. 6. Proof of Theorem 2.2 As a preparatory step we prove the following technical lemma. Lemma 6.1. Let u and v be as in the statement of Theorem 2.2. Suppose that for some k s and some constant η > 0 the following inequality holds
u(k) ∞ η u(k + 1) L∞ (Ω) . (6.1) L (Ω) Then one has u(k + 1)L∞ (Ω) |u(k + 1)| Cη , sup v(k + 1)L∞ (Ω) Ω v(k + 1)
(6.2)
where C is a constant depending only on N, d0 , α0 , r0 , R0 , M0 . For future use we note that for all positive bounded functions a, b on Ω we have the following elementary inequalities a aL∞ (Ω) a (6.3) inf sup . ∞ Ω b bL (Ω) Ω b Proof of Lemma 6.1. For any real valued function f , denote by f+ (f− ) the positive (negative) part of f . As above let u(·, t; s, u0 ) denote the solution of (1.1) on Ω × (s, ∞) with the initial condition u(·, s) = u0 . By uniqueness of
× [k, ∞) solutions of initial value problems, we see that for all (x, t) ∈ Ω u(x, t) = u1 (x, t) − u2 (x, t), where u1 (x, t) = u(x, t; k, u+ (k)), u2 (x, t) = u(x, t; k, u− (k)). Our assumption on u implies that for each t > k the functions u1 (·, t) and u2 (·, t) are equal at some point x(t) ∈ Ω. Since they are nonnegative, by Lemma 3.4 either both of them are identically zero or they are positive in Ω × (k, ∞). The first case is trivial so we shall assume that u1 (and hence also u2 ) is positive. Taking this into account and applying Theorem 2.1, we get sup Ω
u2 (k + 1) u2 (k + 1) C1 inf C1 , Ω u1 (k + 1) u1 (k + 1)
where C1 = C1 (N, α0 , d0 , r0 , R0 , M0 ). Using Theorem 3.3 and our assumption (6.1), we derive
u1 (k + 1) ∞ C u1 (k) L∞ (Ω) = C u+ (k) L∞ (Ω) C u(k) L∞ (Ω) Cη u(k + 1) L∞ (Ω) . L (Ω)
(6.4)
(6.5)
Estimates in (6.3) and Theorem 2.1 imply the following inequalities inf Ω
u1 (k + 1) u1 (k + 1)L∞ (Ω) u1 (k + 1) u1 (k + 1) C1 inf . sup Ω v(k + 1) v(k + 1) v(k + 1)L∞ (Ω) Ω v(k + 1)
Utilizing (6.6), (6.4), and (6.5), we finally get sup Ω
|u(k + 1)| u1 (k + 1) |u(k + 1)| u1 (k + 1) |u(k + 1)| = sup sup sup v(k + 1) Ω v(k + 1) u1 (k + 1) Ω v(k + 1) Ω u1 (k + 1) u1 (k + 1) u1 (k + 1) + u2 (k + 1) u1 (k + 1)L∞ (Ω) sup C1 (1 + C1 ) C1 inf Ω v(k + 1) Ω u1 (k + 1) v(k + 1)L∞ (Ω) u(k + 1)L∞ (Ω) C1 C(1 + C1 )η , v(k + 1)L∞ (Ω)
completing the proof.
2
(6.6)
726
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
We are ready to give the proof of Theorem 2.2. Proof of Theorem 2.2. Let u and v be as in the statement of the theorem. Theorem 3.3 gives
u(t) ∞ C u(s + n) ∞ for t ∈ [s + n, s + n + 1], n 0 L (Ω)
L (Ω)
with a constant C independent of u, s, n. This estimate and Corollary 3.10 imply u(t)L∞ (Ω) u(s + n)L∞ (Ω) C1 ∞ v(t)L (Ω) v(s + n)L∞ (Ω)
(t ∈ [s + n, s + n + 1], n 0)
(6.7)
with a possibly bigger constant C1 , which is again independent of u, v, s, n. Let μ ∈ (0, 1) be as in (2.5) and denote by C2 the constant in Corollary 3.10 (with δ set equal to 1 in it). We have the following two (mutually exclusive) possibilities: (a) For all k s v(k + 1)L∞ (Ω) u(k + 1)L∞ (Ω) μ μ < u(k)L∞ (Ω) C2 v(k)L∞ (Ω) or else (b) the assumptions of Lemma 6.1 are satisfied for some k with η = C2 /μ. Assume (a) holds. Using first (6.7) and afterward the assumed inequality repeatedly, we find that u(s + [t − s])L∞ (Ω) u0 L∞ (Ω) u(t)L∞ (Ω) C1 C1 μ[t−s] v(t)L∞ (Ω) v(s + [t − s]L∞ (Ω) v(s)L∞ (Ω) C1 −(− log μ)(t−s) u0 L∞ (Ω) e , μ v(s)L∞ (Ω)
(t s) (6.8)
where [t] stands for the integer part of t. We thus get the estimate of Theorem 2.2 with C = Assume now that (b) occurs, i.e., for some k s
C1 μ
and γ = − log μ.
u(k + 1)L∞ (Ω) μ . u(k)L∞ (Ω) C2 Let us call k0 the smallest k s such that the above inequality holds. Just as in (a), we get estimate (6.8) for all t ∈ [s, k0 ]. Note that, in view of (6.7), the same estimate holds for t ∈ [s, k0 + 1] possibly after C1 is made larger. Suppose now t k0 + 1. We claim supΩ (|u(t˜ )|/v(t˜ )) < ∞ for all t˜ > s. Indeed, we can write u(x, t˜ ) = u1 (x, t˜ ) − u2 (x, t˜ ), where u1 (x, t˜ ) = u(x, t˜; s, u+ (s)), u2 (x, t˜ ) = u(x, t˜; s, u− (s)) are nonnegative solutions of (1.1). By Theorem 2.1 one has supΩ (ui (t˜ )/v(t˜ )) < ∞, i = 1, 2, which proves the claim. This observation enables us to use Corollary 2.3. Using successively estimates (6.7), (6.3), (2.5), estimate (6.2) with k = k0 and η = Cμ2 , and (6.8) for t ∈ [s, k0 + 1], we deduce u(t)L∞ (Ω) u(k0 + [t − k0 ])L∞ (Ω) |u(k0 + [t − k0 ])| u(k0 + [t − k0 ]) C1 osc C1 C1 sup Ω v(t)L∞ (Ω) v(k0 + [t − k0 ])L∞ (Ω) v(k + [t − k ]) v(k0 + [t − k0 ]) 0 0 Ω u(k0 + 1) |u(k0 + 1)| 2C1 μ[t−k0 ]−1 sup C1 μ[t−k0 ]−1 oscΩ v(k0 + 1) Ω v(k0 + 1) ∞ u(k CC + 1) u(s + [k0 + 1 − s])L∞ (Ω) 2 0 L (Ω) 2C1 μ[t−k0 ]−1 2C12 CC2 μ[t−k0 ]−2 ∞ μ v(k0 + 1)L (Ω) v(s + [k0 + 1 − s])L∞ (Ω) u0 L∞ (Ω) u0 L∞ (Ω) 2C3 2C3 μ[t−k0 ]−2 μ[k0 +1−s] 3 μ(t−s) . v(s)L∞ (Ω) v(s)L∞ (Ω) μ Thus in this case we have established estimate (2.3) from Theorem 2.2 with C = Since all possibilities were covered, the proof is complete. 2
2C3 μ3
and γ = − log μ, as desired.
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
727
Remark 6.2. We remark that in our general setting, the unique continuation theorem may not hold. Therefore it may
the solution u(·, t; s, u0 ) changes sign for all t in some interval (s, t1 ) happen, in principle, that for some u0 ∈ C0 (Ω) and it is identical to 0 for t = t1 (hence, by Theorem 3.3, also for all t t1 ). The statement of Theorem 2.2 remains valid for such initial data as well. Indeed, as in the above proof, one shows that the estimate is valid for t ∈ (s, t1 ). For t t1 , the estimate is trivial. 7. Proofs of Theorem 2.6 and Theorem 2.7 Proof of Theorem 2.6. For the proof of existence of the positive entire solution ϕL , we refer the reader to [20]. The ¯ L ) follows immediately from Corollaries 3.10 and 5.3. This proves statement (i). uniform bound on λ(ϕL ), λ(ϕ We continue with the proof of (iii). Let u0 ∈ XL2 (s) (with XL2 (s) as defined in statement (ii)). Then, either u(·, t; s, u0 ) changes sign for all t, or it becomes (and remains) nonnegative or nonpositive. In the latter case it must be identically zero for all sufficiently large t, for the Harnack inequality rules out the possibility of it being nontrivial and having a zero in Ω. Thus Theorem 2.2 applies to u(·, t; s, u0 ) (see also Remark 6.2). Estimate (2.11) now follows from (2.3) and an easy combination of Corollary 3.10 and the L2 –L∞ estimates from Proposition 5.1. Finally, we prove (ii). The invariance properties of XLi (t), t ∈ R, i = 1, 2, are obvious from the definitions. To prove that XL2 (t) is a closed subspace of X (for XL1 (t) this is trivial), we first give an equivalent characterization of the set XL2 (s), for any s ∈ R. Fix γ as in (iii). Then u(·, t; s, u0 )X γ (t−s) e is bounded for t s . (7.1) XL2 (s) = u0 ∈ X: ϕL (·, t)X Indeed, by statement (iii), the expression in (7.1) is bounded for each u0 ∈ XL2 (s). On the other hand, if u0 ∈ / XL2 (s), then u(·, t; s, u0 ) is of one sign (positive or negative) for all large t. Then Theorem 2.1 with u1 = ϕL , u2 = |u(·, t; s, u0 )| readily implies that the expression in (7.1) is unbounded. From (7.1) and the linear dependence of u(·, t; s, u0 ) on u0 it clearly follows that XL2 (s) is a subspace of X. To prove that it is closed, consider a sequence un ∈ XL2 (s) approaching some u0 ∈ X. It follows from statement (iii) that the expression in (7.1) with u0 replaced by un is bounded by a constant independent of n. Taking the limit we obtain that the expression with u0 is bounded, hence u0 ∈ XL2 (s). It remains to prove (2.10). Obviously, XL1 (t) ∩ XL2 (t) = {0}. Let u0 ∈ X. The arguments in the proof of Theorem 2.3 show that supΩ |u(t˜, t, u0 )|/ϕL (t˜ ) < ∞ for all t˜ > t. As in the proof of Corollary 2.4, for some q ∈ R we have sup Ω
u(t˜, t, u0 ) q, ϕL (t˜)
inf Ω
u(t˜, t, u0 ) q ϕL (t˜)
as t˜ → ∞. Then w(t˜ ) := u(t˜, t, u0 ) − qϕL (t˜ ) is a solution of (1.1) on Ω × (t, ∞) which has a zero in Ω for all t˜ t. In particular, w(t) ∈ XL2 (t). Since u0 = w(t) + qϕL (t), (2.10) is proved. 2 Proof of Theorem 2.7. The proof of the existence of ψL with the required properties is the same as the proof of (i) of Theorem 2.6. To prove the characterization for XL2 (t), t ∈ R, we refer to the following general fact. If u is a solution of (1.1) on (−∞, t0 ) and v is the solution of (2.12) with v(·, t0 ) = v0 ∈ L2 (Ω) then, using (5.4), one easily verifies that we have (7.2) u(·, t), v(·, t) := u(x, t)v(x, t) dx ≡ const. Ω
This relation implies that the orthogonal complement (in L2 (Ω)) of ψL (t) is a codimension one subspace of L2 (Ω) contained in XL2 (t). Therefore it must be equal to XL2 (t). 2 8. Proof of Theorem 2.8 We shall use the same notation as in the statement of the theorem. Let UL (·, ·), UL˜ (·, ·) be the evolution operators ˜ respectively, see Subsection 5.1. They have the following continuity property with respect to the associated with L, L, coefficients.
728
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
Lemma 8.1. Let L, L˜ be as in Theorem 2.8 and assume that (2.14) holds for some δ > 0. Then there is a constant C = C(N, α0 , d0 , R0 ) such that for any t, s ∈ R with 0 t − s 1 we have
UL (t, s) − U ˜ (t, s) 2 (8.1) L L(L (Ω)) Cδ. Proof. Let u0 ∈ L2 (Ω) and let u be the solution of (5.1). Similarly, for the same u0 , let u˜ be the solution of (5.1), ˜ We will use some well known facts from [24] (see Chapter III, in particular) and adhere to where L is replaced by L. the notation used in that book. Let w = u − u. ˜ Then it is the solution of the following initial value problem vt + Lv = ∂i fi − f in Ω × (s, ∞), v = 0 on ∂Ω × (s, ∞), v = 0 in Ω × {s},
(8.2)
where fi = (aij − a˜ ij )∂j u˜ + (ai − a˜ i )u, ˜ ˜ f = (bi − bi )∂i u˜ + (c0 − c˜0 )u˜ (we use the summation convention as above). The energy inequality [24, Theorem III.2.1] implies N
fi L2 (Ω×(s,s+1)) + f L2 (Ω×(s,s+1)) max w(t) L2 (Ω) + wL2 ((s,s+1),W 1,2 (Ω)) C
sts+1
0
i=1
Cδu ˜ L2 ((s,s+1),W 1,2 (Ω)) ,
(8.3)
0
where C depends only on N, α0 , d0 , R0 . Now, the same energy inequality applied to the solution u˜ of (5.1) gives that the right-hand side of (8.3) is bounded above by Cδu0 L2 (Ω) with a possibly larger C. Putting these estimates together, we obtain
max w(t) L2 (Ω) Cδu0 L2 (Ω) . sts+1
Since w(t) = UL (t, s)u0 − UL˜ (t, s)u0 and u0 ∈ L2 (Ω) was arbitrary, the assertion of Lemma 8.1 is proved.
2
for a suitable constant C,
whenever Remark 8.2. Note that by Proposition 5.1 we have UL (t, s)L(L2 (Ω)) C 0 t − s 1. Combining this with estimate (8.1) and the fact that UL (·, ·) satisfies the usual composition property
n−1 δ, where C is as in Lemma 8.1, n 1 is an of evolution operators, we get UL (t, s) − UL˜ (t, s)L(L2 (Ω)) nCC
n−1 is estimated from above by (2 max{C,
C})n , integer and t, s ∈ R are such that 0 t − s n. Obviously, nCC n which, in turn, can be written as C for another constant C depending only on the constants listed in Lemma 8.1. Thus, replacing the assumption 0 t − s 1 in Lemma 8.1 by 0 t − s n, n being an integer with n 1, we get estimate (8.1) with C replaced by C n . We will use this observation in the sequel. We next prove the continuity properties of λ¯ (ϕL ), λ(ϕL ). Proposition 8.3. Let L, L˜ be as in Theorem 2.8. For each ε > 0 there exist numbers C(ε) > 0, δ(ε) > 0, depending only on ε and N , α0 , d0 , r0 , R0 , M0 , such that if (2.14) holds with δ δ(ε) then for all t s ϕ ˜ (t)L∞ (Ω) e−ε(t−s) ϕL˜ (t)L∞ (Ω) ϕL (t)L∞ (Ω) C(ε)eε(t−s) L . C(ε) ϕL˜ (s)L∞ (Ω) ϕL (s)L∞ (Ω) ϕL˜ (s)L∞ (Ω)
(8.4)
Remark 8.4. By Corollary 5.3, the L∞ -norms in Proposition 8.3 can be replaced by the Lp -norms for any p ∈ [1, ∞). Proof of Proposition 8.3. We start with several estimates of quotients appearing in (8.4). Fix s ∈ R and let uL (t) = UL (t, s)1, that is, uL is the solution of (5.1) with the initial condition uL (·, s) ≡ 1. For any t > s define max (t) := supΩ (uL (t)/ϕL (t)), min (t) := infΩ (uL (t)/ϕL (t)). Thus oscΩ (uL (t)/ϕL (t)) = max (t) − min (t). Then Corollary 2.3
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
729
implies that limt→∞ oscΩ (uL (t)/ϕL (t)) = 0 and by Corollary 3.2 we also have that max (t) (min (t)) is a positive nonincreasing (nondecreasing) function of t. Thus for some q(s) > 0 the following holds lim max (t) = lim min (t) = q(s).
t→∞
(8.5)
t→∞
Utilizing (7.2), one easily shows that uL (t) 1ψL (s) dx = lim uL (t)ψL (t) dx = lim ϕL (t)ψL (t) dx = q(s) ϕL (s), ψL (s) . t→∞ t→∞ ϕL (t) Ω
Ω
Ω
Thus q(s) = Ω ψL (s)dx/kL , where kL := ϕL (s), ψL (s) is independent of s by (7.2). By Proposition 5.1, uL (s + 1)L∞ (Ω) C1 for some constant C1 . Equality (8.5) combined with inequalities in (6.3) (applied to a = uL and b = ϕL ), a repeated application of estimate (2.5), Theorem 2.1, the upper bound on uL (s + 1)L∞ (Ω) and Corollary 3.10 imply uL (s + n)L∞ (Ω) oscΩ uL (s + n) μn−1 oscΩ uL (s + 1) μn−1 sup uL (s + 1) − q(s) ϕ (s + n) ∞ ϕL (s + n) ϕL (s + 1) L L (Ω) Ω ϕL (s + 1) ∞ uL (s + 1) uL (s + 1)L (Ω) Cμn−1 inf Cμn−1 Ω ϕL (s + 1) ϕL (s + 1)L∞ (Ω) μn−1
CC1 C 2 C1 μn−1 , ϕL (s + 1)L∞ (Ω) ϕL (s)L∞ (Ω)
where n 1 is an arbitrary integer and μ is as in (2.5). Define
ψL (s) dx cL (s) := ϕL (s) L∞ (Ω) Ω = ϕL (s) L∞ (Ω) q(s). ϕL (s), ψL (s)
(8.6)
(8.7)
Obviously cL (s) 1. Applying Lemma 3.9 to both ϕL and ψL , one finds a constant C2 > 1 such that cL (s) ∈ [1, C2 ] for any s ∈ R. Using (8.6) and (8.7), we obtain
ϕL (s + n)L∞ (Ω) n−1 ϕL (s + n)L∞ (Ω) uL (s + n) ∞ − cL (s) , (8.8) C3 μ L (Ω) ϕ (s) ∞ ϕ (s) ∞ L
L (Ω)
L
L (Ω)
˜ we derive where C3 1 . Carrying out again the same procedure with L replaced in all places by L,
ϕ (s + n)L∞ (Ω) ϕ ˜ (s + n)L∞ (Ω) C3 μn−1 L˜ u ˜ (s + n) ∞ − cL˜ (s) L , L L (Ω) ∞ ϕ (s) ϕ (s) ∞ = C2C
L˜
L (Ω)
L˜
(8.9)
L (Ω)
with obvious meanings of the new notation. Assume now that (2.14) holds for some δ > 0. Using the definitions of uL , uL˜ and Lemma 8.1 (see also Remark 8.2), we get the following bound
uL (s + n) ∞ − uL˜ (s + n) L∞ (Ω) uL˜ (s + n) − uL (s + n) L∞ (Ω) C n δ. (8.10) L (Ω) Obviously, (8.8), (8.9) and (8.10) imply ∞ ϕ (s + n)L∞ (Ω) cL (s) ϕL (s + n)L (Ω) − c ˜ (s) L˜ L ϕL (s)L∞ (Ω) ϕL˜ (s)L∞ (Ω) ϕ ˜ (s + n)L∞ (Ω) ϕL (s + n)L∞ (Ω) C3 μn−1 + C3 μn−1 L + C n δ, ϕL (s)L∞ (Ω) ϕL˜ (s)L∞ (Ω)
(8.11)
where, as before, n 1 is an arbitrary integer and μ ∈ (0, 1) is as in (2.5). Note that we have proved above that for a suitable constant C2 one has 1 cL (s) C2 and the same is true for cL˜ (s), s ∈ R. Choose now an integer n0 1 such that in (8.11) we have C3 μn−1 1/2 for all n n0 (n0 is determined only by N , α0 , d0 , r0 , R0 , M0 ). The choice of n0 , the two sided bounds on cL (s), cL˜ (s), and (8.11) imply that for all n n0 ϕ ˜ (s + n)L∞ (Ω) ϕL (s + n)L∞ (Ω) C4 L + C n δ, ϕL (s)L∞ (Ω) ϕL˜ (s)L∞ (Ω)
(8.12)
730
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
where C4 is some positive constant. Although we fixed s ∈ R at the beginning of this proof, all estimates that we derived are independent of s, hence (8.12) is valid for any s ∈ R. To complete the proof of the proposition, let ε > 0 be arbitrary. Let C5 be the maximum of the constants appearing in (8.12), (3.9), and (5.3), and fix n n0 so large that log(C5 + 1)/n < ε. Set δ = C5−2n , and assume that (2.14) holds. With this choice of δ, (8.12) implies ϕ ˜ (s + n)L∞ (Ω) ϕL (s + n)L∞ (Ω) (C5 + 1) L . (8.13) ∞ ϕL (s)L (Ω) ϕL˜ (s)L∞ (Ω) Now, given any t s, we write t − s = kn + r, where k 0 is an integer and 0 r < n. Then, repeatedly using (3.9), (8.13), and (5.3) (with p = q = ∞), we obtain ϕ ˜ (s + kn)L∞ (Ω) ϕL (t)L∞ (Ω) ϕL (s + kn)L∞ (Ω) C5n C5n (C5 + 1)k L ϕL (s)L∞ (Ω) ϕL (s)L∞ (Ω) ϕL˜ (s)L∞ (Ω) ∞ ϕ ˜ (t)L (Ω) C52n (C5 + 1)k L . ϕL˜ (s)L∞ (Ω) We can rewrite (8.14) as follows ϕ ˜ (t)L∞ (Ω) ϕ ˜ (t)L∞ (Ω) ϕL (t)L∞ (Ω) C52n eC5 (t−s)/n L Ceε(t−s) L ∞ ∞ ϕL (s)L (Ω) ϕL˜ (s)L (Ω) ϕL˜ (s)L∞ (Ω)
(t s),
(8.14)
(8.15)
with C = C52n . Interchanging the role of L and L˜ in the above arguments, starting from (8.11), we obtain (8.15) ˜ These estimates give (8.4). We also see that δ(ε), C(ε) in Proposition 8.3 can be expressed with L replaced with L. −1/ε , C(ε) = C 1/ε , respectively, where C depends only on N , α0 , d0 , r0 , R0 , M0 . Thus the proof of as δ(ε) = C Proposition 8.3 is complete. 2 ¯ L ). Proposition 8.3 and Remark 8.4 imply the statement of Theorem 2.8(i) regarding λ(ϕL ) and λ(ϕ We derive the remaining statements of Theorem 2.8 from well-known robustness properties of exponential dichotomies for abstract evolution operators. We first introduce the concept of exponential dichotomy. By an evolution operator T on a Banach space X we mean a family {T (t, s); t, s ∈ R, t s} ⊂ L(X) with the following two properties (I denotes the identity on X): T (t, s)T (s, r) = T (t, r),
T (t, t) = I
(t, s, r ∈ R, t s r),
t → T (t, s)x: [s, ∞) → X is continuous for all x ∈ X, t, s, ∈ R. Definition 8.5. We say that an evolution operator T admits exponential dichotomy (on R) if there are positive constants , κ and projections P (t) ∈ L(X), t ∈ R, such that (i) T (t, s)P (s) = P (t)T (t, s) (t, s ∈ R, t s); (ii) For each t s, the restriction T (t, s)|R(P (s)) of T (t, s) to the range of P (s) is an isomorphism onto R(P (t)); we define T (s, t) as the inverse of this isomorphism; (iii) T (t, s)(I − P (s))L(X) κe−(t−s) (t, s ∈ R, t s); (iv) T (t, s)P (s)L(X) κe−(s−t) (t, s ∈ R, t < s). The constants and κ (which are of course not unique) are referred to as an exponent and bound of the dichotomy. It is not difficult to prove (cf. Exercise 4 in [16, Section 7.6]) that if T admits exponential dichotomy, then the projections P (t) are uniquely determined (interestingly, a similar uniqueness property for exponential separations is not valid). Given the projections P (t), one defines the associated Green’s function G(t, s) by T (t, s)(I − P (s)) if t > s, G(t, s) = T (t, s)P (s) if t < s. Observe that and κ being an exponent and bound for the dichotomy is equivalent to the Green’s function satisfying the estimate
G(t, s)
κe−|s−t| (t = s). L(X)
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
731
Also note that limt→s+ G(t, s) = I − P (s) in the strong (pointwise) operator topology. In the following proposition we summarize the standard robustness properties of exponential dichotomies. Proposition 8.6. Let T , T be evolution operators on a Banach space X satisfying
K := sup T (t, s) L(X) < ∞, sup T (t, s) − T(t, s) L(X) < ε. t,s∈R 0t−s1
t,s∈R t−s=1
Assume T admits exponential dichotomy with exponent , bound κ, and projections P (t). Given any ˜ < , κ˜ > κ, there exists ε1 depending only on κ, κ, ˜ , , ˜ and K, with the following property. If ε < ε1 , then T admits exponential (t) satisfying dichotomy with exponent , ˜ bound κ, ˜ and projections P
P (t) − P (t)
(8.16) L(X) < Cε (t ∈ R), where C is a constant depending only on κ, , and K. Proof. With the exception of (8.16), the statement is the same as that of Theorem 7.6.10 in [16]. Estimate (8.16) s) − G(t, s)L(X) , for the associated Green’s functions, as given in the proof follows from the estimate on G(t, of [16, Theorem 7.6.10]. Note that the proofs of these results in [16] are based on the relation between discrete and continuous dichotomies and apply to abstract evolution operators, independently of any underlying parabolic equation. 2 To apply the above abstract result, we define the following evolution operators on X = L2 (Ω) U,L (t, s) := e(t−s) V,L˜ (t, s) := e(t−s)
ϕL (s)L2 (Ω) ϕL (t)L2 (Ω) ϕL (s)L2 (Ω) ϕL (t)L2 (Ω)
(8.17)
UL (t, s), UL˜ (t, s)
(t s).
(8.18)
Here UL and UL˜ are the evolution operators of (1.1) and (2.12), respectively, and is a suitable positive number. A key observation employed below is that if L admits exponential separation with bound C and exponent γ , and ∈ (0, γ /2] , then U,L admits exponential dichotomy with exponent and some bound κ. Indeed, define projections PL (t) by PL (t)u =
u, ψL (t) ϕL (t) ϕL (t), ψL (t)
(u ∈ X).
(8.19)
Note that PL (t) has the range and kernel given by the invariant subspaces XL1 (t), XL2 (t) of Theorem 2.6. Also note that PL (t) is bounded: by Corollary 5.3, for some C0 > 0 ϕL (t) ψL (t) C0 1, (8.20) , ϕL (t)L2 (Ω) ψL (t)L2 (Ω) which gives the bound
−1
PL (t)
L(X) C0 .
(8.21)
Using this bound, properties of XL1 (t), XL2 (t), and (2.11), it is straightforward to verify that U,L admits exponential dichotomy with exponent , bound C(1 + C0−1 ) and projections PL (t). ˜ then the projection P ˜ (t) corresponding Building on this observation, our first aim is to show that if L is close to L, L ˜ to L is close to PL (t) and, consequently, statement (i) of Theorem 2.8 holds. Lemma 8.7. There exist constants δ0 and C depending only on N , α0 , d0 , r0 , R0 , M0 , such that if (2.14) holds with δ δ1 , then
PL (t) − P ˜ (t) 2 (8.22) L L(L (Ω)) Cδ (t ∈ R)
732
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
and
ϕL˜ (t) ϕL (t)
− Cδ
ϕ (t) 2 ϕL˜ (t)L2 (Ω) L2 (Ω) L L (Ω)
(t ∈ R).
(8.23)
Proof. Let γ = γ (N, α0 , d0 , r0 , R0 , M0 ) be as in Theorem 2.6 and set = γ /2. As remarked above, U,L admits exponential dichotomy with exponent and projections PL (t). Suppose (2.14) holds with δ > 0 sufficiently small (as specified below). Using Corollary 5.3 and Lemma 8.1, one gets
eγ /2 Cδ = C1 δ (8.24) sup U,L (t, s) − V ˜ (t, s) 2 ,L
t−s=1 s,t∈R
L(L (Ω))
and also, by Proposition 5.1 and Corollary 5.3,
sup U,L (t, s)
= K < ∞,
(8.25)
L(X)
t,s∈R 0t−s1
where C1 and K depend only on N , α0 , d0 , r0 , R0 , M0 . By Proposition 8.6, if δ is small enough, say δ < δ˜ = ˜ (t) ˜ δ(N, α0 , d0 , r0 , R0 , M0 ), then V,L˜ admits exponential dichotomy with exponent γ /4 and some projections P L satisfying
PL (t) − P ˜ (t) 2 < C2 δ (t ∈ R), L
L(L (Ω))
with C2 = C2 (N, α0 , d0 , r0 , R0 , M0 ). Now assume also that δ δ(γ /8), where δ(γ /8) is given by Proposition 8.3 (with ε = γ /8). Then, as one easily verifies, V,L˜ at the same time admits exponential dichotomy with exponent γ /4 and projections PL˜ (t). As remarked above, the projections of exponential dichotomies are uniquely determined. Thus ˜ (t), which proves (8.22). PL˜ (t) = P L In the remaining part of this proof · means · L2 (Ω) . By (8.22),
ϕL (t)
ϕL (t) ϕL (t)
ϕL (t)
= − PL˜ (t) − PL˜ (t) Cδ PL (t) ϕL (t) ϕL (t) ϕL (t) ϕL (t)
ϕL (t) ϕL˜ (t) ϕL˜ (t) ϕL (t)
. − (8.26)
− − PL˜ (t) ϕL (t) ϕL˜ (t) ϕL˜ (t) ϕL (t)
We obviously have ϕL˜ (t) ϕL (t) = PL˜ (t) . ϕL˜ (t) PL˜ (t)ϕL (t) Thus
ϕL˜ (t) PL˜ (t)ϕL (t) ϕL (t) ϕL (t)
ϕL (t)
ϕ (t) − PL˜ (t) ϕ (t) = 1 − ϕ (t) ϕ (t) − PL˜ (t) ϕ (t) Cδ. L L L L ˜ L
Comparing (8.26) and (8.27), we conclude
ϕL˜ (t) ϕL (t)
−
ϕ (t) 2 ϕ (t) 2 2 L
L (Ω)
L˜
2Cδ
(t ∈ R).
(8.27)
2
L (Ω) L (Ω)
We can now complete the proof of Theorem 2.8(i). By Corollary 5.3 and Proposition 5.2, the function ϕL˜ (t) ϕL (t) − ϕL (t)L2 (Ω) ϕL˜ (t)L2 (Ω)
by a constant depending on N , α0 , d0 , r0 , R0 , M0 . is bounded in L∞ (Ω) and consequently in a Hölder space C α (Ω), ∞ This and (8.23) imply that the L -norm of this function tends to 0 as δ → 0 (uniformly in t). Finally, we note that the normalizations in L2 (Ω) can be replaced by the normalizations in L∞ (Ω). Indeed, the previous L∞ -estimate and Corollary 5.3 imply that ϕL˜ (t)L2 (Ω) ϕL (t)L2 (Ω) − ϕL (t)L∞ (Ω) ϕL˜ (t)L∞ (Ω)
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
733
tends to 0 as δ → 0. Estimating the function ϕL (t)L2 (Ω) ϕL˜ (t)L2 (Ω) ϕL˜ (t) ϕL˜ (t) ϕL (t) ϕL (t) − = − ϕL (t)L∞ (Ω) ϕL˜ (t)L∞ (Ω) ϕL (t)L2 (Ω) ϕL (t)L∞ (Ω) ϕL˜ (t)L2 (Ω) ϕL˜ (t)L∞ (Ω) using the triangle inequality we obtain the desired conclusion. We are finished with the proof of statement (i) of Theorem 2.8 regarding ϕL and ϕL˜ . The conclusion regarding ψL , ψL˜ is obtained by applying the proved statement to the operators L∗ and L˜ ∗ (see the discussion following Theorem 5.2). We next prove Theorem 2.8(ii). Lemma 8.8. Suppose that L admits exponential separation bound CL and exponent γL . For each ε > 0 there is δ3 = δ3 (ε, CL , γL , N, α0 , d0 , r0 , R0 , M0 ) > 0 such that if (2.14) holds with δ δ3 , then L˜ admits exponential separation with some bound C(ε, CL , γL ) > 0 and exponent γL˜ γL − ε. Proof. Initially, we proceed as in the proof of Lemma 8.7. Suppose that (2.14) holds with δ > 0 sufficiently small (as specified below). Setting = γL /2, we observe that U,L admits exponential dichotomy with exponent γL /2, bound CL (1 + C0−1 ) (cp. (8.21)), and projections PL (t), and that (8.24), (8.25) hold, where the constants C1 and K now also depend on γL . Using Proposition 8.6, we find δ3 = δ3 (ε, CL , γL , N, α0 , d0 , r0 , R0 , M0 ) > 0 such that if δ < δ3 , ˜ (t). These projections are then V,L˜ admits exponential dichotomy with exponent (γL − ε)/2 and some projections P L close to PL (t), in particular, their range is one-dimensional. We make δ3 smaller, if necessary, so that (8.28) δ3 < min δ(ε/2), δ(γ /4) , where δ(·) is given by Proposition 8.3 and γ is as in Theorem 2.6. We claim that ˜ (t) = P ˜ (t). P L
L
(8.29)
Although this time we do not know a priori that V,L˜ admits exponential dichotomy with the latter projections (hence we cannot refer to the uniqueness), we do know, as in the proof of Lemma 8.7, that Vγ /2,L˜ (t, s) = e
γ −γL 2 (t−s)
V,L˜ (t, s)
admits such a dichotomy. This is sufficient for (8.29). For example, by definition of exponential dichotomy, the range ˜ (s) consists of all v ∈ X such that e(γL /2−ε/2)(t−s) V ˜ (t, s)vX is bounded as t → ∞. If ε < 2γL − γ , of I − P ,L L which we may assume without loss of generality (note that, by Theorem 2.6, Lemma 8.8 is trivial if γL γ ), then for any such v also Vγ /2,L˜ (t, s)vX is bounded. Hence, again by definition of exponential dichotomy, v is in the range ˜ (s). These two inclusions of I − PL˜ (s). Similarly one proves that the range of PL˜ (s) is contained in the range of P L give (8.29). Using (8.29), the exponential dichotomy for V,L˜ , and Proposition 8.3 (recall (8.28)), a straightforward estimate implies the conclusion of the lemma. 2 Lemma 8.8 implies statement (ii) of Theorem 2.8. The proof of the theorem is now complete. 9. Proof of Theorem 2.10 This proof will be carried out in several steps analogous to those in Section 5. We work with a fixed operator L as in (D), whose coefficients satisfy (1.4) and (1.2) on RN +1 . Since in this part we consider varying domains, we will henceforth denote by UΩ (·, ·) the evolution operator associated with problem (1.1). It has the following continuity property with respect to the domain. Lemma 9.1. Let Ω1 , Ω2 , with Ω1 ⊂ Ω2 , be any two Lipschitz domains, whose Lipschitz constants satisfy (1.7). Then for each ε > 0 there exists δ(ε) > 0 depending only on ε and N , α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω1 , ∂Ω2 ) δ(ε), then
UΩ (s + 1, s) − UΩ (s + 1, s) ∞ N ε (s ∈ R). (9.1) 1
2
L(L (R ))
734
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
Remark 9.2. As a matter of definition, the evolution operator UΩ (t, s) acts on L∞ (Ω), but it has a natural extension to L∞ (RN ). Specifically, given any u0 ∈ L∞ (RN ), we first compute UΩ (t, s)(u0 |Ω ), where u0 |Ω is the restriction of u0 to Ω and then extend the resulting function by zero outside Ω. This way, UΩ (t, s) can be viewed as a continuous operator on L∞ (RN ). We use this convention throughout the section and, abusing the notation slightly, we use the same symbol UΩ (t, s) for the extended operator. Proof of Lemma 9.1. Given any constant δ > 0 and a domain Ω, we set Ω δ = x ∈ Ω: dist(x, ∂Ω) > δ . For each σ > 0 (small enough) let ζσ ∈ C0∞ (RN ), 0 ζσ 1, be a function with ζσ ≡ 1 on Ω12σ , ζσ ≡ 0 on RN \ Ω1σ and ∇ζσ L∞ (Ω1 ) C/σ for some constant C. We can always choose ζσ so that the constant C depends only on N , r0 , R0 , and M0 . Let ui , i = 1, 2, be the solutions of (1.1) on Ωi × (s, ∞), respectively, such that ui (x, s) = ζσ (x) in Ωi , i = 1, 2, where s ∈ R is fixed from now on. Note that u1 , u2 are positive. If we define w = u1 − u2 , then w satisfies (weakly) wt + Lw = 0 in Ω1 × (s, ∞) with w(·, s) ≡ 0 on Ω1 and w(x, t) = u2 (x, t) for (x, t) ∈ ∂Ω1 × (s, ∞). Using a standard L∞ estimate (see [24, Chapter III, Theorem 7.1]), we get sup
Ω1 ×[s,s+1]
|u1 − u2 | C
sup
∂Ω1 ×[s,s+1]
(9.2)
u2 ,
where C depends only on N, α0 , d0 , R0 . Theorem 3.3 guarantees that u2 is bounded on Ω2 × [s, s + 1] by a constant depending only on N , α0 , d0 . This fact, our choice of ζσ and Theorem 5.2 imply that α u2 C(σ ) d(∂Ω1 , ∂Ω2 ) (9.3) sup ∂Ω1 ×[s,s+1]
for some α > 0, which depends only on N, d0 , α0 . The constant C(σ ) depends only on σ , α, N , α0 , d0 , r0 , R0 , M0 . Combining the last two inequalities, we obtain α (9.4) sup |u1 − u2 | C(σ ) d(∂Ω1 , ∂Ω2 ) . Ω1 ×[s,s+1]
By Proposition 5.4, we can write ui (x, t) = ki (x, t; ξ, s)ζσ (ξ ) dξ
((x, t) ∈ Ωi × (s, ∞)),
(9.5)
Ωi
where ki (x, t; ξ, s) is the weak Green’s function associated with problem (1.1) with Ω replaced by Ωi , i = 1, 2. Statements (ii), (iii) in Proposition 5.4 in combination with Theorem 5.2 imply that we have the following inequality α ki (x, s + 1; ξ, s) C d(∂Ω1 , ∂Ω2 ) + 2σ (x ∈ Ωi \ Ω12σ , ξ ∈ Ωi , i = 1, 2), (9.6) with α > 0 the same as in (9.4) and C depends only on α, N , α0 , d0 , r0 , R0 , M0 . Using (iv) in Proposition 5.4, the choice of ζσ , (9.5), (9.4) and (9.6), we derive
UΩ (s + 1, s) − UΩ (s + 1, s) ∞ N = sup k2 (x, s + 1; ξ, s) − k1 (x, s + 1; ξ, s) dξ 1
L(L (R ))
2
k2 (x, s + 1; ξ, s) − k1 (x, s + 1; ξ, s) dξ
sup x∈Ω2 Ω2 \Ω12σ
+
sup x∈Ω2 \Ω12σ
+ sup x∈Ω12σ
x∈Ω2 Ω2
k2 (x, s + 1; ξ, s) − k1 (x, s + 1; ξ, s) dξ
Ω12σ
k2 (x, s + 1; ξ, s) − k1 (x, s + 1; ξ, s) dξ
Ω12σ
α C Ω2 \ Ω 2σ + C Ω 2σ d(∂Ω1 , ∂Ω2 ) + 2σ + sup u1 (s + 1) − u2 (s + 1) 1
1
Ω12σ
α α C Ω2 \ Ω 2σ + C1 d(∂Ω1 , ∂Ω2 ) + 2σ + C(σ ) d(∂Ω1 , ∂Ω2 ) , 1
(9.7)
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
735
where |Ω| denotes the measure of Ω, the constants C, C1 , C(σ ) depend on α, N , α0 , d0 , r0 , R0 , M0 and C(σ ) in addition depends on σ . Notice that, under our assumptions on Ω1 , Ω2 , the measure |Ω2 \ Ω12σ | can be estimated from above by C(d(∂Ω1 , ∂Ω2 ) + 2σ ), where C depends only on N , r0 , R0 , M0 . Hence,
UΩ (s + 1, s) − UΩ (s + 1, s) ∞ N C d(∂Ω1 , ∂Ω2 ) + 2σ α + C(σ ) d(∂Ω1 , ∂Ω2 ) α . (9.8) 1 2 L(L (R )) Here, we assumed without loss of generality that σ , d(∂Ω1 , ∂Ω2 ) are sufficiently small and α ∈ (0, 1). It is now apparent that, given any ε > 0, it is sufficient to choose σ of order ε1/α and then d(∂Ω1 , ∂Ω2 ) of order (ε/C(σ ))1/α to get the assertion of Lemma 9.1. 2 In the next corollary we remove the restriction Ω2 ⊂ Ω1 . Corollary 9.3. Let Ω1 be a Lipschitz domain, whose Lipschitz constants satisfy (1.7) and let Ω2 be any bounded Lipschitz domain. Then for each ε > 0 there exists δ(ε) > 0, depending only on ε and N , α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω1 , ∂Ω2 ) δ(ε), then
UΩ (s + 1, s) − UΩ (s + 1, s) ∞ N ε (s ∈ R). (9.9) 1 2 L(L (R )) More precisely, there exists δ1 = δ1 (N, α0 , d0 , r0 , R0 , M0 ) such that (9.8) holds, provided d(∂Ω1 , ∂Ω2 ) δ1 and σ δ1 . 1 , Ω 2 such that Proof. Assume d(∂Ω1 , ∂Ω2 ) is small compared to r0 in (1.7). We can then find two domains Ω 1 , ∂ Ω 2 ) 2d(∂Ω1 , ∂Ω2 ), Ω 1 ⊂ Ωi ⊂ Ω 2 , i = 1, 2, and Ω 1 , Ω 2 are Lipschitz domains satisfying (1.5), (1.6), d(∂ Ω replacing the constants r0 , M0 , R0 by r0 /2, 2M0 , 2R0 , if necessary. The result now follows from the monotonicity of the weak Green’s functions with respect to the domain (see Proposition 5.4) and the estimates of Lemma 9.1. 2 Remark 9.4. A similar extension as in Remark 8.2 applies here. Using the boundedness of the evolution operator (Theorem 3.3) one shows that the statements of Corollary 9.3 remain valid if one replaces s + 1 by s + n, n 1 being any integer; in this case, the constants C, C(σ ) in (9.8) have to be replaced by C n , C(σ )n , where C, C(σ ) may have to be made larger, but they do not depend on n ∈ N. With the above estimates on the evolution operators, the proof of Theorem 2.10 can be carried out along the lines of the proof of Theorem 2.8. We limit the further exposition to indicating the necessary modifications. The next proposition is analogous to Proposition 8.3. be as in Theorem 2.10. Then for each ε > 0 there exist numbers C(ε), δ(ε) > 0, depending Proposition 9.5. Let Ω, Ω < δ(ε), then for all t s only on ε and N , α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω, ∂ Ω) ϕΩ (t)L∞ (Ω) e−ε(t−s) ϕΩ (t)L∞ (Ω) ϕΩ (t)L∞ (Ω) C(ε)eε(t−s) . C(ε) ϕΩ (s)L∞ (Ω) ϕΩ (s)L∞ (Ω) ϕΩ (s)L∞ (Ω)
(9.10)
Proof. The only noteworthy difference from the proof of Proposition 8.3. is that we use (9.8) (and Remark 9.4) in place of (8.1) (and Remark 8.2). Thus the estimate corresponding to (8.12) reads as follows ϕΩ (s + n)L∞ (Ω) ϕΩ (s + n)L∞ (Ω) + 2σ α + C(σ ) n d(∂Ω, ∂ Ω) α. C4 + C n d(∂Ω, ∂ Ω) ∞ ϕΩ (s)L (Ω) ϕΩ (s)L∞ (Ω)
(9.11)
One now easily derives the estimate corresponding to (8.13) by first choosing σ sufficiently small followed by a The rest of the arguments used in the proof of Proposition 8.3 are straightforward to suitable choice of d(∂Ω, ∂ Ω). modify. We omit the details. 2 By Corollary 5.3, we can replace in (9.10) the L∞ -norms on the respective domains by the L2 -norms. This implies ¯ Ω ). the statement of Theorem 2.10(i) regarding λ(ϕΩ ) and λ(ϕ
736
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
In the following it is convenient to view the functions ϕΩ (t), ψΩ (t) as elements of Y := L∞ (RN ) (extending them by 0 outside Ω). Understanding that ·, · now stands for the standard inner product in L2 (RN ), we define a continuous projection PΩ (t) on Y by PΩ (t)u =
u, ψΩ (t) ϕΩ (t). ϕΩ (t), ψΩ (t)
(9.12)
With these conventions and notation, arguing similarly as in the proof of Lemma 8.7, one proves the following proposition. be as in Theorem 2.10. Then for each ε > 0 there exists δ(ε) > 0, depending only on ε and Proposition 9.6. Let Ω, Ω δ(ε), then N , α0 , d0 , r0 , R0 , M0 , such that if d(∂Ω, ∂ Ω)
PΩ (t) − P (t) ∞ N ε (t ∈ R), (9.13) Ω
and
L(L (R ))
ϕΩ (t) ϕΩ (t)
− ε
ϕ (t) ∞
ϕΩ (t)L∞ (Ω) L∞ (RN ) Ω L (Ω)
(t ∈ R).
(9.14)
With this result we have completed the proof of the statement of Theorem 2.10(i) regarding ϕΩ , ϕΩ . The conclusion regarding ψΩ , ψΩ is again taken care of by the discussion following Theorem 5.2. The proof of statement (ii) of Theorem 2.10 is completely analogous to the proof of Lemma 8.8 and is omitted. Appendix A Proof of Proposition 5.5. By a standard modification (multiplication of u by e−m(t−s) ), one reduces the proof to the case with m = supΩ×(s,T ) (−c0 ) = 0. Henceforth we assume this extra condition. Let D(Ω) denote the space of smooth functions with compact support in Ω. The proof consist of a two step approximation procedure. First, assuming the initial condition is in D(Ω), we find the solution as the limit of solutions of approximating problems on smooth subdomains of Ω. Then we approximate a general initial condition by functions in D(Ω) and take the limit of the solutions of these approximate problems. We start with preliminary estimates of solutions on approximating subdomains. Fix any 0 ∈ (0, r0 /4], where r0 is as in (1.7). Choose a family of smooth domains Ωε , ε ∈ (0, 0 ), which is decreasing in ε (with respect to inclusion)
ε ⊂ Ω, d(∂Ωε , ∂Ω) ε, and (1.7) is satisfied with Ω replaced by Ωε , and and such that for each ε one has Ωε ⊂ Ω r0 , M0 replaced by r0 /2, 2M0 , respectively. Let fε ∈ D(Ω) be any real-valued function with dist(x, ∂Ω) 20
(x ∈ supp fε )
and consider the following problem ut + Lu = 0 in Ωε × (s, T ), u = 0 on ∂Ωε × (s, T ), u = fε in Ωε × {s}.
(A.1)
ε × [s, T ]) for all p > 1 (see It has a unique solution uε and the solution is contained in W 2,1,p (Ωε × (s, T )) ∩ C(Ω [26, Theorem 7.17]). We claim that for all ∈ (0, 0 ] one has 2 θ0 ε uε (x, t) sup fε (x) ω () := sup (t ∈ [s, T ]), (A.2) 0 x∈Ωε dist(x,∂Ωε ) where θ0 > 0 depends only on N , d0 , α0 , r0 , R0 , M0 . Note that, by the maximum principle, sup uε (x, t) sup fε (x) (t ∈ [s, T ]). x∈Ωε
(A.3)
x∈Ωε
This implies that in order to verify (A.2) it suffices to show that ωε () 2−θ0 ωε (2) for any fixed ∈ [0, 0 /2]. For this purpose, let (x0 , t0 ) be any point in Ωε × [s, T ] with dist(x0 , ∂Ωε ) . The assumptions on Ωε imply that there
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
737
are two positive numbers K1 , K2 , depending only on N , r0 , R0 , M0 , and a point y0 ∈ RN \ Ωε such that the following holds. If we set r = dist(x0 , ∂Ωε )/K1 , then |x0 − y0 | = K2 r and Br (y0 ) ∩ Ωε = ∅. Following ideas of [25], we define the slant cylinder V by (t − (t0 − r 2 )) 2 < r, t (A.4) V := (x, t) ∈ RN × R: x − y0 − (x − y ) − r < t < t 0 0 0 0 . r2 We will show at the end of this proof that there is a function v ∈ C ∞ (RN +1 ) such that v vanishes on the lateral side of ∂V , 0 v 1, vt + Lv 0 in V , and v(x0 , t0 ) κ0 > 0, where κ0 ∈ (0, 1) is a constant depending only on N , d0 , α0 , r0 , R0 , M0 . Assume for now such a function v exists. In W := (Ωε × (s, T )) ∩ V define the function w := uε + ωε (2)v. Clearly, wt + Lw 0 in W . Moreover, since the base of V is outside Ω × (s, T ), properties of v imply that w ωε (2) on the parabolic boundary ∂p W of W . Thus, by the maximum principle, w ωε (2) in W . In particular, w(x0 , t0 ) ωε (2) and hence uε (x0 , t0 ) 1 − v(x0 , t0 ) ωε (2) (1 − κ0 )ωε (2). (A.5) This inequality holds for −u as well and since (x0 , t0 ) was arbitrary we conclude that ωε () 2−θ0 ωε (2) with θ0 = − log2 (1 − κ0 ). Thus estimate (A.2) is established. Consider now an arbitrary f˜ ∈ D(Ω) and set 0 := inf dist(x, ∂Ω): x ∈ supp f˜ ∪ {r0 /4} . For ε ∈ (0, 0 /2) take fε := f˜|Ωε in (A.1). Extending the corresponding solution uε by zero outside Ωε , we view it as a function on Ω. Let 0 < ε1 < ε2 < 0 /2 be arbitrary. Recall that Ωε2 ⊂ Ωε1 , by the monotonicity of the family. The maximum principle and (A.2) imply sup uε (x, t) + sup uε (x, t) sup uε (x, t) − uε (x, t) x∈Ω stT
1
2
x∈Ωε1 \Ωε2 stT
1
x∈∂Ωε2 stT
1
2ε2 θ0 2ε2 θ0 ˜ 2 sup fε1 (x) = 2 sup f (x) . 0 0 x∈Ωε1 x∈Ω
× [s, T ]) and, by the interior Lp -estimates This estimate implies that uε , ε ∈ (0, 0 /2), is a Cauchy family in C(Ω 2,1,p [26, Theorem 7.13], also in (the Fréchet space) Wloc (Ω × (s, T )) for each p ∈ (1, ∞). It follows that, as ε → 0, uε 2,1,p converges to a solution u˜ of (5.1) with u0 = f˜ and the solution is in Wloc (Ω × (s, T )) for each p ∈ (1, ∞). By the 2,1,N +1
× [s, T ])) is uniquely determined (Ω × (s, T )) ∩ C(Ω maximum principle, the solution (as a function in Wloc ˜ by f . We have thus proved Proposition 5.5 under the extra assumption u0 = f˜ ∈ D(Ω). To remove this assumption, take
and choose a sequence fn ∈ D(Ω) such that fn → f in C0 (Ω).
For the solution un of (5.1) an arbitrary f ∈ C0 (Ω) with u0 = fn we have, by the maximum principle, sup um (x, t) − un (x, t) sup fm (x) − fn (x) (m, n = 1, 2, . . .). x∈Ω stT
x∈Ω
× [s, T ]) and in W 2,1,p (Ω × (s, T )) This and the interior Lp -estimates imply that un is a Cauchy sequence in C(Ω loc for each p ∈ (1, ∞). Arguing as above, we conclude that un converges to a unique solution of (5.1) with u0 = f and the solution has the regularity as stated in Proposition 5.5. To finish this proof, we still need to construct the function v with the properties stated above. Define 2 2 2 (t − (t0 − r 2 )) 1 −β (t−(t02−r )) 2 r r − x − y 0 − (x0 − y0 ) , (A.6) v(x, t) := 4 e 2 r r where (x, t) ∈ V and β > 0 is to be determined. It is obvious that 0 v 1 in V and v vanishes on the lateral side of ∂V . Denoting (t − (t0 − r 2 )) A := x − y0 − (x0 − y0 ) and B := r 2 − A2 , 2 r
738
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
by a simple computation, using (ND), (1.4) and (1.2), one shows 2 1 −β (t−(t02−r )) 4N β r 8α0 A2 − 4N d0 B − 4N d0 AB − d0 B 2 − 2 |x0 − y0 |AB + 2 B 2 . −vt − Lv 4 e r r r
(A.7)
From now on, we assume r 1 without loss of generality. Let us first consider the case when A (1 − ε)r and ε > 0 is small. Using this assumption, the fact that A r, |x0 − y0 | = K2 r, r 1 and (A.7), we compute 2 1 −β (t−(t02−r )) 2 r e (r ) 8α0 (1 − ε)2 − 16N d0 ε − 4d0 ε 2 − 8N K2 ε . 4 r The right-hand side of this inequality is nonnegative, independently of β > 0, provided ε is chosen sufficiently small. The choice of ε depends only on N , d0 , α0 and K2 . Fix such a number and call it ε0 . If, on the other hand, A (1 − ε0 )r, an elementary inequality implies that ε0 r 2 B r 2 1. Using this fact and (A.7), we obtain
−vt − Lv
2
1 −β (t−(t02−r )) r e B(−8N d0 − d0 − 4N K2 + βε0 ). r4 If β is big enough, as determined by N , d0 , K2 , ε0 , the right hand side of this inequality becomes positive. In view of the dependences of K2 and ε0 , β can be chosen depending only on N , d0 , α0 , r0 , R0 , M0 . With this β we have vt + Lv 0 in V and v(x0 , t0 ) = e−β . Defining κ0 = e−β , the function v has all the properties claimed above. The proof is now complete. 2 −vt − Lv
References [1] D.G. Aronson, Non-negative solutions of linear parabolic equations, Ann. Scuola Norm. Sup. Pisa 22 (1968) 607–694. [2] J.M. Arrieta, Elliptic equations, principal eigenvalue and dependence on the domain, Comm. Partial Differential Equations 21 (1996) 971–991. [3] H. Berestycki, L. Nirenberg, S.R.S. Varadhan, The principal eigenvalue and maximum principle for second-order elliptic operators in general domains, Comm. Pure Appl. Math. 47 (1994) 47–92. [4] I. Birindelli, Hopf’s lemma and anti-maximum principle in general domains, J. Differential Equations 119 (2) (1995) 450–472. [5] S.-N. Chow, K. Lu, J. Mallet-Paret, Floquet bundles for scalar parabolic equations, Arch. Rational Mech. Anal. 129 (1995) 245–304. [6] D. Daners, Dirichlet problems on varying domains, J. Differential Equations 188 (2003) 591–624. [7] D. Daners, Domain perturbation for linear and nonlinear parabolic equations, J. Differential Equations 129 (1996) 358–402. [8] D. Daners, Existence and perturbation of principal eigenvalues for a periodic-parabolic problem, Electron. J. Differential Equations, Conf. 05 (2000) 51–67. [9] D. Daners, Heat kernel estimates for operators with boundary conditions, Math. Nachr. 217 (2000) 13–41. [10] L.C. Evans, Partial Differential Equations, Graduate Studies in Mathematics, vol. 19, American Mathematical Society, Providence, RI, 1998. [11] E.B. Fabes, N. Garofalo, S. Salsa, A backward Harnack inequality and Fatou theorem for nonnegative solutions of parabolic equations, Illinois J. Math. 30 (1986) 536–565. [12] E.B. Fabes, M.V. Safonov, Behavior near the boundary of positive solutions of second order parabolic equations, J. Fourier Anal. Appl. 3 (1997) 871–882. [13] E.B. Fabes, M.V. Safonov, Y. Yuan, Behavior near the boundary of positive solutions of second order parabolic equations II, Trans. Amer. Math. Soc. 12 (1999) 4947–4961. [14] E. Ferretti, M.V. Safonov, Growth theorems and Harnack inequality for second order parabolic equations, in: Harmonic Analysis and Boundary Value Problems, in: Contemp. Math., vol. 277, Amer. Math. Soc., Providence, RI, 2001, pp. 87–112. [15] N. Garofalo, Second order parabolic equations in nonvariational form: boundary Harnack principle and comparison theorems for nonnegative solutions, Ann. Mat. Pura Appl. 138 (1984) 267–296. [16] P. Henry, Geometric Theory of Semilinear Parabolic Equations, Lecture Notes in Mathematics, vol. 840, Springer, New York, 1981. [17] P. Hess, Periodic-Parabolic Boundary Value Problems and Positivity, Longman Scientific & Technical, Harlow, 1991. [18] P. Hess, P. Poláˇcik, Boundedness of prime periods of stable cycles and convergence to fixed points in discrete monotone dynamical systems, SIAM J. Math. Anal. 24 (1993) 1312–1330. [19] J. Húska, Harnack inequality and exponential separation for oblique derivative problems on Lipschitz domains, J. Differential Equations 226 (2006) 541–557. [20] J. Húska, P. Poláˇcik, The principal Floquet bundle and exponential separation for linear parabolic equations, J. Dynam. Differential Equations 24 (2004) 1312–1330. [21] J. Húska, P. Poláˇcik, M.V. Safonov, Principal eigenvalues, spectral gaps and exponential separation between positive and sign-changing solutions of parabolic equations, in: Disc. Cont. Dynamical Systems, Supplement, Proceedings of the 5th International Conference on Dynamical Systems and Differential Equations, Pomona 2004, 2005, pp. 427–435. [22] V. Hutson, W. Shen, G.T. Vickers, Estimates for the principal spectrum point for certain time-dependent parabolic operators, Proc. Amer. Math. Soc. 129 (6) (2001) 1669–1679 (electronic). [23] N.V. Krylov, M.V. Safonov, A property of the solutions of parabolic equations with measurable coefficients, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1) (1980) 161–175.
J. Húska et al. / Ann. I. H. Poincaré – AN 24 (2007) 711–739
739
[24] O.A. Ladyzhenskaya, V.A. Solonnikov, N.N. Uralceva, Linear and Quasilinear Equations of Parabolic Type, Translation of Mathematical Monographs, American Mathematical Society, Providence, RI, 1968. [25] E.M. Landis, Second Order Equations of Elliptic and Parabolic Type, Translation of Mathematical Monographs, American Mathematical Society, Providence, RI, 1998. [26] G.M. Lieberman, Second Order Parabolic Differential Equations, World Scientific Publishing Co. Inc., River Edge, NJ, 1996. [27] J. Mierczy´nski, Flows on order bundles, unpublished. [28] J. Mierczy´nski, p-arcs in strongly monotone discrete-time dynamical systems, Differential Integral Equations 7 (1994) 1473–1494. [29] J. Mierczy´nski, Globally positive solutions of linear PDEs of second order with Robin boundary conditions, J. Math. Anal. Appl. 209 (1997) 47–59. [30] J. Mierczy´nski, Globally positive solutions of linear parabolic partial differential equations of second order with Dirichlet boundary conditions, J. Math. Anal. Appl. 226 (1998) 326–347. [31] J. Mierczy´nski, The principal spectrum for linear nonautonomous parabolic pdes of second order: Basic properties, J. Differential Equations 168 (2000) 453–476. [32] J. Mierczy´nski, W. Shen, Exponential separation and principal Lyapunov exponent/spectrum for random/nonautonomous parabolic equations, J. Differential Equations 191 (2003) 175–205. [33] J. Moser, A Harnack inequality for parabolic differential equations, Comm. Pure Appl. Math. 17 (1964) 101–134; Correction in Comm. Pure Appl. Math. 20 (1967) 231–236. [34] M. Nishio, The uniqueness of positive solutions of parabolic equations of divergence form on an unbounded domain, Nagoya Math. J. 130 (1993) 111–121. [35] P. Poláˇcik, Parabolic equations: asymptotic behavior and dynamics on invariant manifolds, in: B. Fiedler (Ed.), Handbook on Dynamical Systems, vol. 2, Elsevier, Amsterdam, 2002, pp. 835–883. [36] P. Poláˇcik, On uniqueness of positive entire solutions and other properties of linear parabolic equations, Discrete Contin. Dynamical Systems 12 (2005) 13–26. [37] P. Poláˇcik, I. Terešˇcák, Convergence to cycles as a typical asymptotic behavior in smooth strongly monotone discrete-time dynamical systems, Arch. Rational Mech. Anal. 116 (1992) 339–360. [38] P. Poláˇcik, I. Terešˇcák, Exponential separation and invariant bundles for maps in ordered Banach spaces with applications to parabolic equations, J. Dynamics Differential Equations 5 (1993) 279–303; Erratum, J. Dynamics Differential Equations 6 (1) (1994) 245–246. [39] D. Ruelle, Analycity properties of the characteristic exponents of random matrix products, Adv. in Math. 32 (1979) 68–80. [40] W. Shen, Y. Yi, Almost automorphic and almost periodic dynamics in skew-product semiflows, Mem. Amer. Math. Soc. 647 (1998) 93. [41] I. Terešˇcák, Dynamics of C 1 smooth strongly monotone discrete-time dynamical systems, Preprint. [42] I. Terešˇcák, Dynamical systems with discrete Lyapunov functionals, PhD thesis, Comenius University, 1994.